Sie sind auf Seite 1von 434

SYSTEM RELIABILITY OF FIXED OFFSHORE STRUCTURES

UNDER FATIGUE DETERIORATION

A thesis submitted for the degree of


Doctor of Philosophy in the faculty of Engineering
of the University of London

by

Navilkumar Shetty

Department of Civil Engineering


Imperial College of Science, Technology and Medicine
LONDON

April 1992

UNW
Dedicated to the memory of my late mother Pushpavati Sheny
111

ABSTRACT

The focus of research on the reliability of offshore structures has so far been mainly
on the reliability analysis of the structure under extreme wave loading. Most of the
reported studies are, however, inherently limited to idealized member behaviour with
simplified treatment of uncertainty in environmental loading where the effect of
fatigue damage at tubular joints is not accounted for. In this thesis a more rigorous
methodology has been presented for system reliability analysis of a jacket structure
under both extreme wave and fatigue conditions.

A fully consistent, comprehensive and yet reasonably simple mechanical model has
first been developed for the assessment of tubular joints. A short-term and a
long-term probability density function model for hot-spot stress-ranges have been
developed to compute cumulative fatigue loading accurately even for broad-banded
stress spectra. A fracture mechanics fatigue crack propagation model has been
developed to estimate crack growth through the thickness and at the surface of a
tubular joint which explicitly accounts for the effects of weld geometry, residual
stresses, crack coalescence, variable amplitude loading and fatigue threshold. An
approach has also been proposed for using the CEGB R6 method for fracture
assessment of tubular joints. Reliability analysis of isolated tubular joints has been
carried out for the limit-states of fatigue failure, plastic collapse and fatigue-fracture.

A new system reliability method for the extreme wave condition of a jacket structure
has been developed, jointly with others within the CEC BR1TE Project No. P 1270.
The method is based on the joint beta-point concept for the reliability formulation and
a virtual distortion method for progressive collapse analysis. An efficient response
surface model is used for the treatment of uncertainties in environmental loading.
This methodology is extended to include fatigue effects in the form of deterioration
in the capacity of tubular joints. A new method of failure-tree enumeration has been
proposed which uses selective enumeration and preferential branching to identify
failure sequences in the order of their importance so as to compute bounds on system
reliability to within user-specified limits.
iv
A general formulation for time-dependent system reliability has been developed for
analysis of a sequence of pure fatigue failures or a combination of fatigue and fracture
failures. Finally, an approach for combining both time-dependent and extreme load
failures has been proposed. This method allows the development of a single
failure-tree for the structure which includes: (i) pure extreme load failure sequences
in which all elements fail under a single large wave, (ii) pure fatigue/fracture failure
sequences in which brittle elemental failures occur at random points in time during the
service of the structure and (iii) mixed failure sequences in which initial failures occur
by fatigue and/or fracture at random points in time and the significantly weakened
structure then collapses under a single large wave.

The proposed system reliability methodology has been combined with associated
developments within the CEC BR1TE project so that a fully integrated software
package called RASOS has now become available for practical application of system
reliability concepts in the design and assessment of offshore jacket structures. The
methodology and the software has been used for the reliability analysis of a 6-legged
jacket platform, based on which a number of conclusions have been drawn. The
proposed method, however, is limited for application to jacket structures in shallow
to moderate water depths for which dynamic effects are not significant.
V

ACKNOWLEDGEMENTS

My foremost thanks ale to my supervisor Prof. Michael J Baker, Dept. of Engineering,


University of Aberdeen (formerly at Imperial College) for his help and guidance
throughout the course of my studies. His painstaking study of the manuscript of the
thesis, constructive comments and suggestions have greatly helped in improving the
presentation and this is deeply appreciated.

My Ph.D study and stay in the U.K. has been funded by the Commonwealth
Scholarship Commission of the U.K. and this is gratefully acknowledged. I would
like to thank my employers Oil and Natural Gas Commission for granting study leave
and the Govt. of India for nominating me for the scholarship. Additional funding for
research has come from Elf Aquitaine, CTICM and WS Atkins Engineering Sciences
which is also gratefully acknowledged.

The work described in the thesis has been carried out within the framework of the
European Commission BRITE Project No. P 1270 and has considerably benefitted
from the associated developments in the project. The co-operation of all the project
partners and sponsors is acknowledged. In particular I would like to acknowledge the
discussions with Prof. P Thoft-Christensen, Prof. A Vrouwenvelder, Dr iT Gierlinski,
Mr RC Turner, Dr S Gollwitzer, Dr Y Guenard, Mr J Goyet, Dr B Mariem,
Dr J Labeyrie, Dr M Birades and Mr J Hello. I would also like to thank the BR1TE
Project Steering Committee and the RASOS Users' Group for their permission to use
the RASOS software for solving some of the examples presented in the thesis.

I would like to thank Prof. PJ Dowling, Head of Civil Engineering Dept. and
Prof. R Hobbs, Head of Structures section for their administrative support. I would
also like to thank Mrs. S Wright for typing part of the thesis, Ms. F Donovan and
Mr G Scopes for all their help during my stay at Imperial College.
vi
My special thanks are to Mr V Balakiishnan for his keen interest in reading the
manuscript and to Mr RJ Sears for providing computer drawings of the structural
model studied in the thesis.

My heartfelt thanks are to nmbers of the staff Dr K Ramachandran,


Dr MK Chryssanthopoulos, Dr PC Davidson and to my colleagues and friends Marc,
Frederick, Gill, Dion, Fadi, Harsha and others in the department who have helped n
to enjoy my long stay at Imperial College.

My deepest appreciation and thanks are to my wife Sumangala who has withstood the
pressures of being a research student's wife and has always inspired my work, and to
my new born son Nitin whose arrival expedited the completion of the thesis. Finally
my thanks are to my parents, brother, sisters and my in-laws for their well wishes and
encouragement and for patiently waiting for my return home.
vu

CONTENTS


1. INTRODUCTION 1-1

1.1 GENERAL 1-1

1.2 NATURE OF WADING AND RESPONSE OF OFFSHORE STRUCTURES 1-2

13 BEHAVIOUR OF PRIMARY MEMBERS AND JOINTS 1-3

1.4 ROLE OF REDUNDANCY AND SYSTEM BEHAVIOUR 1-5

1.5 ROLE OF INSPECTION AND MAINThNANCE PROGRAMS 1-6

1.6 CURRENT DETERMINISTIC DESIGN PRACTICE 1-8

1.7 NEED FOR PROBABILITY BASED DESIGN AND ASSESSMENT METHODS 1-12

1.8 OBJECTIVES OF THE THESIS 1-14

1.9 ORGANISATION OF THE THESIS 1-18

2. A CRITICAL REVIEW OF COMPUTATIONAL METHODS FOR



STRUCTURAL SYSTEM RELIABILITY ANALYSIS 2-1

2.1 GENERAL 2-1

2.2 PROBABILISTIC BASIS OF STRUCTURAL SAFETY 2-2

2.3 TREATMENT OF UNCERTAINTIES N A RELIABILITY ANALYSIS 2-4

2.4 ELEMENT RELIABILITY ANALYSIS 2-5

2.4.1 Problem fonnulation 2-6

2.4.2 Second moment reliability indices 2-7

2.4.3 Evaluation of failure probability using FORM and SORM 2-12

2.4.3.1 First ordez reliability method 2-14

2.4.3.2 Second order reliability method 2-19

2.5 USEFUL PROBABILITY TRANSFORMATIONS 2-20

2.5.1 Normal correlated variables 2-20

2.5.2 Non-normal uncorrelated variables 2-21

2.5.3 Non-normal correlated variables 2-22

2.5.4 Non-normal marginals with known correlation matrix 2-23

2.6 SYSTEM RELIABILITY ANALYSIS 2-25

2.6.1 Reliability analysis of parallel systems 2-25

2.6.1.1 First order reliability method for parallel systems 2-26

2.6.1.2 Second order reliability method for parallel systems 2-27

2.6.13 Probability bounds for parallel systems 2-31
viu

2.6.2 Reliability analysis of series systems 2-32

2.6.2.1 First order reliability method for series systems 2-32

2.6.2.2 Second order reliability method for series systems 2-33

2.6.2.3 Probability bounds for series systems 2-33

2.6.3 Reliability analysis of general systems 2-35

2.7 SENSITIVITY AND IMPORTANCE MEASURES 2 37

2.7.1 Direction cosine sensitivity factors 2-37

2.7.2 Omission sensitivity factors 2-38

2.7.3 Parametric sensitivity factors 2-40

2.8 CONCLUDING REMARKS 2-43

3. MODELS FOR LOADING AND RESPONSE OF OFFSHORE STRUCTURES 3-1



3.1 GENERAL 3-1

3.2 PROBABILISTIC DESCRIPTION OF OCEAN ENVIRONMENT 3-2

3.2.1 Short-term wave climate modelling 3-3

3.2.2 Long-term wave climaie modelling 3-6

3.2.3 Extreme wave condition modelling 3-8

3.2.4 Joint occurrence of wave, wind and current conditions 3-10

3.3 METHODS OF RESPONSE ANALYSIS 3-11

3.3.1 Deterministic, discrete wave static analysis 3-12

3.3.2 Spectral dynamic analysis 3-13

3.4 PROBABILISTIC DESCRIPTION OF STRUCTURAL RESPONSE 3-16

3.4.1 Short-term response modelling 3-16

3.4.2 Long-term response modelling 3-19

3.4.3 Extreme response modelling 3-20

3.5 PROBABILISTIC MODELLING OF STRESS CYCLES FOR FATIGUE 3-22

3.5.1 Review of available methods 3-23

3.5.1.1 Methods based on deterministic response analysis 3-24

3.5.1.2 Methods based on spectral response analysis 3-25

3.5.2 Proposed model for short-term distribution of stress-ranges 3-27

3.5.2.1 Probability density function for stress-ranges 3-27

33.2.2 Expression for cumulative fatigue loading 3-29

3.5.23 Comparison with previous models 3-32

3 5.3 Proposed model for long term distribution of stress-ranges 3-35

3.5.3.1 Method of moments 3-36

3.5.3.2 Method of Chi-square minimization 3.37

3.5.4 Incorporation of stress threshold effects 3-41
ix

4. MECHANICAL MODELS FOR CAPACITY OF TUBULAR JOINTS 4-1

4.1 GENERAL 4-1

4.2 PLASTIC COLLAPSE OF TUBULAR JOINTS 4-6

4.2.1 Compression loaded joints 4-4

4.2.2 Tension loaded joints 4-7

4.2.3 In-plane moment loaded joints 4-8

4.2.4 Out-of-plane moment loaded joints 4-9

4.2.5 Influence of chord loads on joint strength 4-10

4.2.6 Strength of cracked joints 4-12

4.2.7 Joints under combined loading 4-14

4.3 STRESS DISTRIBUTIONS IN TUBULAR JOINTS 4-15

4.3.1 Geometric Stresses 4-16

4.3.1.1 Stress concentration factors 4-17

4.3.1.2 Stress distribution around the intersection 4-19

4.3.13 Through-thickness stress distribution 4-20

4.3.2 Notch stresses 4-21

4.3.3 Residual stresses 4-23

4.4 FATIGUE FAILURE OF TUBULAR JOINTS 4-26

4.4.1 Fatigue characteristics of offshore structures 4-26

4.4.1.1 Factors affecting fatigue behaviour 4-26

4.4.1.2 Observed crack propagation behaviour in tubular joints 4-28

4.4.2 Fatigue analysis using S-N approach 4-29

4.4.2.1 Basic S-N curve 4-30

4.4.2.2 Effect of plate thickness 4-31

4.4.2.3 Effect of sea water environment 4-33

4.4.2.4 Effect of weld improvement techniques 4-33

4.4.2.5 Effect of variable amplitude loading 4-34

4.4.2.6 Computation of fatigue damage 4-36

4.4.2.7 Comments on the use of S-N approach 4-36

4.4.3 Fatigue analysis using fracture mechanics approach 4-38

4.4.3.1 Review of crack propagation models 4-39

4.4.3.2 Review of stress intensity factor solutions 4-46

4.4.4 Proposed methodology for fatigue analysis of tubular joints 4-50

4.4.4.1 Stress intensity factors 4-51

4.4.4.2 Crack propagation model 4-55
4.4.43 Computation of fatigue damage 4-70
4.4.4.4 Comparison with experimental results 4-73
x

4.5 FRACTURE FAILURE OF TUBULAR JOINTS 4-77

4.5.1 Review of available methods for fracture assessment 4-77

4.5.1.1 Energy balance approach 4-78

4.5.1.2 Stress intensity factor approach 4-79

4.5.1.3 Crack tip opening displacement approach 4-84

4.5.1.4 J integral approach 4-85

4.5.1.5 BS PD 6493 Approach 4-88

4.5.1.6 CEGB R6 Approach 4-92

4.5.2 Proposed methodology for fracture assessment of tubular joints 4-93
4.5.2.1 Failure assessment diagram

4-94

4.5.2.2 Plastic collapse parameter L 4-95

4.5.2.3 Fracture parameter K 4-97
4.5.2.4 Material fracture toughness K. 4-98
4.5.2.5 Comparison with experimental results 4-101


5. RELIABILITY ANALYSIS OF TUBULAR JOINTS 5-1

5.1 GENERAL 5-1

5.2 UNCERTAINTIES IN EXTREME WAVE LOADING 5-2
5.2.1 Environmental description 5-3

5.2.2 Calculation of hydrodynamic loading 5-8
5.2.3 Global structural analysis

5-11

5.3 UNCERTAINTIES IN FATIGUE LOADING 5-11

5.3.1 Explicit consideration of all loading variables 5-12

53.1.1 Long-term wave climate modelling 5-12

53.1.2 Short-term wave climate modelling 5-14

5.3.1.3 Response transfer function 5-15

5.3.2 Uncertainties modelled through parameters of the long-term Weibull model 5-17
5.3.3 Uncertainties modelled through a single variable 5-18

5.4 UNCERTAINTIES IN STRENGTH PREDICTION MODELS 5-19

5.4.1 Plastic collapse load model 5-19

5.42 Stress analysis model 5-21

5.4.3 Fatigue crack propagation model 5-22

5.4.4 Fracture assessment model 5-23

5.5 UNCERTAINTIES IN MATERIAL PROPERTIES 5 24

5.5.1 Elastic material properties 5-24
55.2 Fatigue material properties 5-25
5.5.3 Fracture material properties 5 27
xi

5.6 RELIABILITY ANALYSIS FOR THE LIMIT-STATE OF FATIGUE 5-29

5.6.1 Reliability analysis using the S-N approach 5-29

Example 1: Loading uncertainties modelled through a single variable 5-32

Example 2: Loading uncertainties in terms of Weibull parameters 5-35

5.6.2 Reliability analysis using a fracture mechanics approach 5-38

Example 3: Variation of reliability index with service exposure 5-39

5.7 RELIABILITY ANALYSIS FOR THE LIMIT-STATE OF PLASTIC COLLAPSE 5-43

Example 4: Analysis of an intact tubular joint 5-44

5.8 RELIABILITY ANALYSIS FOR THE LIMIT-STATE OF FRACTURE 5-47

5.8.1 Reliability analysis based on CEGB R6 fracture assessment model 5-47

Example 5: Reliability assessment for the limit-state of fracture 5-51

5.8.2 Reliability analysis based on BS PD 6493 fracture assessment model 5-55

Example 6: Reliability assessment of a weld defect using BS PD 6493 model 5-55

5.9 CONCLUDING REMARKS 5-57

6. SYSTEM RELIABILITY ANALYSIS UNDER AN EXTREME WAVE EVENT 6-1



6.1 GENERAL 6-1

6.2 LIMIT AND POST-LIMIT BEHAVIOUR OF MEMBERS AND JOINTS 6-3

6.2.1 Models for tubular members 6-3

6.2.2 Models for tubular joints 6-8

6.3 STRUCTURAL COLLAPSE ANALYSIS 6-10

6.3.1 Direct plastic mechanism approach 6-10

6.3.2 Successive elastic analyses approach 6-12

6.3.3 Non-linear progressive collapse analysis approach 6-14

6.4 RESPONSE SURFACE APPROACH FOR ENVIRONMENTAL WADING 6-19

6.5 RELIABILITY ANALYSIS FOR A PRESCRIBED FAILURE MODE 6-24

63.1 Reliability analysis based on plastic mechanism approach 6-25

6.5.2 Reliability analysis based on member replacement approach 6-27

6.5.3 Reliability analysis based on joint beta-point concept 6-31

6.5.4 Modified VDM algorithm for prescribed failure sequence analysis 6-34

6.6 SYSTEM RELIABILITY ANALYSIS 6-39

6.6.1 Identification of dominant collapse mechanisms 6-40

6.6.2 Failure-tree enumeration methods 6-43

6.6.3 A practical approach to failure-tree enumeration 6-46

6.7 RELIABILITY ANALYSIS UNDER FATIGUE DETERIORATION 6-52

6.7.1 Incorporation of plastic collapse and fracture of tubular joints 6-54

6.7.2 Incorporation of fatigue deterioration 6-55
xii

6.7.3 Choice of failure elements included in failure-tree enumeration 6-56

Example 1: Reliability analysis of a jacket platform under extreme wave 6-58

Preliminary linear elastic analysis 6-59

Deterministic progressive collapse analysis 6-61

System reliability analysis 6-64

Probabilistic modelling 6-64

Idealization of member behaviour 6-65

Selection of failure elements 6-65

Definition of system failure 6-66

Choice of tnincaiion criterion 6-66

Overview of failure-tree enumeration 6-67

Discussion of results 6-69

Example 2: Reliability analysis under fatigue deterioration 6-77

6.8 CONCLUDING REMARKS 6-79


7. TIME-DEPENDENT SYSTEM RELIABIUTY ANALYSIS 7-1

7.1 GENERAL 7-2

7.2 REVIEW OF AVAILABLE TIME-DEPENDENT RELIABILITY METHODS 7-2

7.3 TIME-DEPENDENT SYSTEM RELIABILiTY FORMULATION 7-6

7.3.1 Time-dependent prescribed failure sequence 7-6

7.3.2 Time-dependent failure-tree enumeration 7-11

7.3.3 Calculation of failure probabilities 7-11

7.4 ANALYSIS OF A SEQUENCE OF FATIGUE FAiLURES 7-12

7.4.1 Approximate method for deriving the damage ratio y 7-12

7.4.2 Analysis using an S-N approach 7-14

7.4.3 Analysis using a fracture mechanics approach 7-16

7.5 ANALYSIS OF A SEQUENCE OF FATIGUE AND FRACTURE FAILURES 7-17

7.6 ANALYSIS OF COMBINED EXTREME LOAD AND FATIGUE FAILURES 7-19

Example 1: System reliability of a jacket structure under fatigue condition 7-22

7.7 CONCLUDING REMARKS 7-27


8. CONCLUSIONS AND RECOMMENDATIONS 8-1

8.1 GENERAL 81

8.2 IMPORTANT CONTRIBUTIONS 82

8.3 SUMMARY OF FINDINGS 8-6

8.4 POTENTIAL APPUCATIONS 8-10

8.5 SUGGESTIONS FOR FURTHER WORK 8-11
xl"

REFERENCES R-1

APPENDICES

A OUTLINE OF THE RASOS SOFTWARE A-i

B LIMTF-STATh TYPES CONSIDERED N RASOS B-I

C BASIC VARIABLES CONSIDERED IN RASOS C-I

D BRITE JACKET PLATFORM D-1
xiv
xv

NOTATION

a0 initial crack depth


a(t) crack depth after time t
a critical crack depth at failure
weld toe plastic zone size
a/c crack aspect ratio
aTF relative crack depth
API American Petroleum Institute
B active stress influence matrix
B1 model uncertainty in collapse load calculation
Bd,b model uncertainty in degree-of-bending factors
B1 model uncertainty in extreme environmental load calculation
B5 model uncertainty in global response transfer function
Bf model uncertainty in stress concentration factors
Bf model uncertainty in stress intensity factors
B .S. British Standards
C Paris law coefficient
Co initial crack semi-length
c(t) crack semi-length after time t
Cd drag coefficient
Cm inertia coefficient
CEGB Central Electricity Generating Board, U.K.
Coy coefficient of variation
CTOD Crack Tip Openeing Displacement
D chord diameter,
Miner's fatigue damage indicator;
stress influence matrix
d brace diameter
base damage rate in joint j corresponding to stress S,
the damage rate corresponding to stress S after (i-i) failures
xvi
d diameter of crack tip plastic zone
da/dn fatigue crack growth rate
d water depth
E[X] expected value of varaible X
Fx(X) cumulative distribution function of variable X
probability density function of variable X
F ultimate tensile strength
F yield stress
FORM First-Order Reliability Method
0 strain energy release rate
G(.) threshold factor
g gap between two braces of a tubular joint
g(.) limit-state function
limit-state function in standard normal space
H wave height
H significant wave height
extreme wave height in a reference time of 'r
J J-integral (fracture parameter)
K material constant of S-N curve;
stress intensity factor
k wave number
Kc plane stress fracture toughness
plane strain fracture toughness
K02 fracture toughness corresponding to 0.2mm crack extension
stress intensity factor range
Alçff effective stress intensity factor range
'Snax maximum stress intensity factor in a loading cycle
minimum stress intensity factor in a loading cycle
Kc,p crack opening stress intensity factor

AISh threshold stress intensity factor range


K5 geometric stress concentration factor
KmE material fracture toughness
xvii

K notch stress concentration factor

Kr non-dimensional fracture parameter
limit value of Kr for fracture failure

L chord length
Lr non-dimensional collapse parameter

M safety margin (limit-state function value)

M applied moment (in- or out-of-plane) in the brace

M plastic limit moment (in- or out-of-plane) of the joint
m slope of da/dn curve; inverse slope of S-N curve

N number of stress cycles to failure (S-N analysis)

n(t) number of stress cycles in time t

P(.) probability of event (.)
Pa applied axial load in the brace
axial plastic limit load of the joint
Pu
probability of failure
pf

q elliptical integral of the second kind

R stress ratio; stress intensity factor ratio;
fracture resistance to crack extension

S hot-spot stress range
SB fatigue strength obtained from the basic S-N curve
Sb bending component of the hot-spot stress
Si base stress in joint j corresponding to intact state of the structure
(i)
the stress level in joint j during th time interval after (i-i) failures
Sm membrane component of the hot-spot stress
max hot-spot stress due to the extreme wave
Sn notch stress
Sop crack opening stress
Sr non-dimensional plastic collapse parameter
residual stress
threshold stress range
So stress endurance limit of S-N curve
SORM Second-Order Reliability Method
xviii
T plate thickness; chord plate thickness;
wave period
T base fatigue life of a joint j corresponding to stress S
T zero-crossing period of waves
t service exposure; reference time of analysis;
brace plate thickness
t) actual time-to-failure of joint j following (j-i) failures
tB reference plate thickness (S-N analysis)
tmg thickness of marine growth
U standard normal variable;
effective stress intensity factor range ratio
U beta-point or design point in standard normal space
U water particle velocity
U water particle acceleration
vc current speed

vx coefficient of variation of variable X


VDM Virtual Distortion Method of non-linear analysis
x a basic random variable
x outcome of random variable X
Y a normal correlated variable;
normalized stress;
stress intensity geometry correction factor
Ycxp stress intensity compliance function from experiments

Yw weld geometry stress intensity correction factor

a sensitivity factor (direction cosine);


joint geometry parameter (2Ld'D);
spectral irregularity factor
reliability index
joint geometry parameter (d/D)
A Miner's damage sum at failure
3 truncation criterion for a failure sequence
xix

&app applied crack tip displacement



6mat critical crack tip displacement (fracture resistance)
8r non-dimensional fracture parameter
limit value of 8r for fracture failure
crack tip opening displacement

C strain;
spectral band-width parameter

CL strain from linear-elastic analysis

Cv virtual strain
virtual distortion

T(.) gamma function
joint geometry parameter (D/2T)
incomplete gamma function
(i) ratio of base damage rate in joint j to the damage rate
during the 1th time interval,

C joint gap parameter (gfD)
stress gradient correction factor

C(a) residual stress relaxation factor

0 angle of inclination of the brace with the chord
random multipliers for response to environmental loading
mean value of variable X

V0 frequency of zero-crossings

VP frequency of peaks
thickness correction exponent (S-N analysis)

(a) bending stress relaxation factor

P degree-of-bending factor
plasticity correction factor
PC

Pu correlation coefficient between X1 and X

a stress
stress from linear-elastic analysis
virtual stress
standard deviation of variable X
xx

joint geomeiry parameter (uT);



(i)
remaining life of joint j after 1,2,...,(i-1) joints have failed in that sequence
n-dimensional standard normal distribution function
n-dimensional standard normal density function
fatigue loading function
fatigue resistance function

(1) frequency of stress cycles
Chapter 1
INTRODUCTION

1.1 GENERAL

Since the discovery of oil off the coast of the Gulf of Mexico in the early fifties many
structures have been designed and constructed to support drilling and production
facilities in the sea. These structures are of several types which can be broadly
categorised as floating structures such as ships and semi-submersibles, fixed structures
such as steel jacket structures and concrete gravity platforms and the more recent
compliant structures such as tension leg platforms and guyed towers. But, by far the
commonest type of structures have been the jacket-type fixed platforms and these are
likely to continue to be used for oil and gas production in moderate water depths for
the foreseeable future. Owing to the high risk and high cost of construction and
operation the safety of these structures is of paramount importance and is the subject
of the present thesis.

A jacket-type fixed platform essentially consists of a deck superstructure which houses


the main drilling and production facilities and accommodation modules, a steel
template or the jacket structure which, besides supporting the top deck, has to resist
the action of ocean waves and current. The jacket structure itself is fixed to the sea
bed by a number of piles driven through or around the main legs of the jacket.

Jacket-type offshore structures are quite distinct from land-based structures in two
main respects which are often a major source of difficulty and high cost in the design,
construction and operation of these structures. Firstly, the uncertainty associated with
the prediction of design environmental factors such as storm wave height, wave period
and current velocity and the determination of environmental forces on the structure.
Secondly, jacket structures are fabricated using cylindrical steel members as they give
the benefit of having a low drag coefficient, high buoyancy, high strength to weight
ratio and equal bending strength in all directions. But, the tubular welded joints at the
ends of members exhibit a complex mode of behaviour and often form the most
1-2
critical parts of the structure. Unavoidable welding defects, residual stresses and the
cyclic action of waves combined with the corrosive environment of the ocean cause
fatigue cracks to initiate and grow at the tubular joints, necessitating regular
under-water inspection and maintenance of these structures. To insure against the
uncertainty of extreme environmental forces to which the structure will be subjected
in its lifetime and to compensate for the high cost of in-service inspection and
maintenance, a considerable degree of redundancy is built into the design of such
structures. For the proper design of jacket structures a good understanding of these
elements is necessary. An overview of the important aspects involved in the design
of offshore structures is presented in the following sections before discussing the need
for probability based methods and formally stating the objectives of the present thesis.

1.2 NATURE OF LOADING AND RESPONSE OF OFFSHORE STRUCTURES

The loading on offshore structures can be considered to be mainly from two sources.
The first is that induced by the gravity and arising from permanent and
semi-permanent equipment, stores items, human occupancy and operational loads,
together with the self-weight of the jacket and deck structures. Most of these loads
do not vary appreciably with time and can be predicted reasonably accurately. The
second source of loading is that caused by the environment and comprises mainly the
forces due to wind on the equipment, deck structures, and the exposed parts of the
jacket, the action of ocean waves and current on the jacket and in some cases ice
forces and earthquake loading. These forces vary appreciably with time and it is
generally difficult to predict the set of extreme environmental conditions to which the
structure will be subjected during its lifetime.

For the in-place analysis and design of the intact structure the first step is to predict
the wave height, wave penod, current velocity, mean wind velocity and gust factor
corresponding to the most extreme storm. However, for fatigue damage assessment
a detailed description of the variation of the sea conditions over the entire life of the
structure is required. Such information is obtained from detailed oceanographic
1-3
measurements at the offshore location over several years prior to the design of the
structure and the use of statistical forecasting techniques.

The next step involves the calculation of environmental forces on various members
and response analysis of the structure due to all loads. For the extreme loading
condition and for platforms in relatively shallow water, wind and wave action is
calculated as a static force applied to the structure and only the static response of the
structure is evaluated. However, for taller structures, for which the natural period of
vibration approaches the period of the ocean waves, a more complicated dynamic
analysis is necessary. In this case it is also not possible to choose, apriori, a set of
environmental conditions which will cause the extreme response in the structure.

Fatigue damage in offshore structures is mainly due to the oscillatory nature of ocean
waves and this damage usually occurs at the welded tubular joints in the structure.
Fatigue damage is governed by the cyclic stress reversal and the stress-range
magnitude at the crack location in a joint as a wave passes through the structure.
Since every wave may cause some damage, in principle, the total number of stress
cycles and their stress-range magnitude due to all waves passing through the structure
during its lifetime is required. Since this is hardly practicable the long-term wave
climate at the offshore location is described by a discrete number of "representative"
sea conditions or "sea-states" and the response of the structure in each of these
sea-states is evaluated using spectral analysis techniques.

1.3 BEHAVIOUR OF PRIMARY MEMBERS AND JOINTS

The jacket structure of an offshore platform is a 3-dimensional steel frame consisting


of 4, 6 or 8 main leg members which are interconnected by several levels of
horizontal and inclined bracing members. The brace members provide lateral
resistance to the structure against wave and current forces. The leg members, besides
supporting the top deck, also act as 'sleeves' for the piles to which they are connected
by injecting cement grout in the annulus between the pile and the sleeve. Leg
1-4
members thus have to resist the vertical loads from the deck and horizontal forces
transferred from the brace members. The pile members support the total vertical loads
and horizontal shear and over-turning moment transferred from the structure at the
level of the sea bed.

The ultimate strength of the tubular bracing members depends on the loading mode.
Under axial tension or bending moment full section plasticity will be mobilized and
under increasing load "plastic hinges" may develop at the ends of the member and at
mid-span. Significant strain-hardening could take place before ultimate capacity is
achieved. Beyond this point plastic strain increases rapidly and after some limit may
even cause "rupture" of the member. Under compressive load the buckling behaviour
of the member is strongly influenced by "initial out-of-straightness" and presence of
any "dents" which often result from boat impacts or dropped objects from the deck.
The behaviour of brace members is also strongly influenced by the tubular welded
joints at the ends of the member. These joints offer considerable flexibility and often
exhibit non-linear moment-rotation characteristics.

Primary members of the jacket structures are inter-connected by welding the profiled
end of one member onto the undisturbed exterior of another member. The member
with the larger diameter and thickness is usually called a "chord" member. Because
of circular cross-sections of connecting members, the tubular joint so formed results
in a complicated geometry. This geometry, along with the presence of the weld,
induces major distortions of the nominal stress field in the brace member. The
difference in axial stiffness of the brace and bending stiffness of the chord wall results
in severe stress concentrations. The presence of the weld increases this concentration
even further. The stress distribution at the joint thus varies both in the thickness
direction and around the intersection.

The loading on the joint is primarily that transferred through the brace member The
brace loads are resisted at the joint by the "shell bending" of the chord tube. Thus the
load-deformation characteristics and load-carrying capacity of the tubular joint will be
very different from those of the brace and chord members. Under static loading, the
1-5

welded joint can fail by local buckling of the chord wall under a compressive brace
load while for a tensile load it will fail by excessive plastic deformation of the chord.
Excessive distortion of the joint is observed under applied moments.

Since tubular joints are formed by welding, small weld defects are invariably present.
Because of the high stress concentrations and cyclic action of large number of waves
the weld defects may grow by fatigue to form long cracks. Furthermore, the corrosive
action of the sea water considerably increases this fatigue damage. If left unrepaired
a fatigue crack may continue to grow through the thickness of the member and along
the welded intersection and the joint may eventually fail by the separation of brace
from the chord member. This type of failure under cyclic loads is called failure by
fatigue and is one of the most important localised failure modes observed in offshore
structures. Fatigue is a particularly serious problem for North Sea structures which
are subject to frequent storms.

The presence of cracks alters the ultimate behaviour of the joint considerably. In
addition to the plastic (or collapse) type of failures a cracked joint can also fail by
fracture under extreme loading. The presence of welding residual stresses will further
increase the risk of fracture. Failure by fracture is affected by several factors such as
overall geometry of the joint, the local geometry in the vicinity of the crack, size and
shape of the crack, weld size and profile and fracture toughness of the parent material,
the weld and the heat affected zone etc. For fracture assessment the detailed stress
pattern at the tip of the crack and stress variation along the crack propagation path
should be known.

1.4 ROLE OF REDUNDANCY AND SYSTEM BEHAVIOUR

Structural redundancy implies a configuration in which the failure of a single element


does not lead immediately to catastrophic total collapse. A redundant structure will
be able to transmit forces through alternative load paths even after the "failure" of one
or more structural elements. A considerable degree of redundancy is often built into
1-6

the design of jacket structures. The need for redundancy arises for several reasons;
(i) the risk of the actual environmental loads exceeding the design loads during normal
platform life, (ii) the unavoidable nature of certain fabrication defects such as
misalignments, out-of-roundness of tubulars, welding defects at the joints, (iii)
accidental damage during service such as due to boat impacts or dropped objects and
(iv) deterioration during service as a result of corrosion and fatigue or fracture failure
initiating from defects or cracks during overload conditions.

The ultimate capacity of the structure depends on the post-limit behaviour of the
primary members and joints, the structural topology and the degree of redundancy.
The primary members can support increasing loads beyond first yield under axial
tension or moment loads due to plastic deformation and strain-hardening, while under
axial compression peak capacity is maintained for only a limited range of deflection
after which the member unloads due to buckling. Similarly, tubular joints under axial
tension reach a peak load-carrying capacity of almost twice the load corresponding to
first yield, but loading beyond this point results in separation of the brace from the
chord due to rupture and complete loss of the load carrying capacity of the joint and
members. Thus it is important to describe accurately the post-limit behaviour of
members and joints in the ultimate strength assessment of the structure.

For rational assessment of the safety of the structure it is necessary to treat the
behaviour of the entire structure as a system and relate structural safety to certain
system limit states. The system limit state could be either the collapse of the entire
structure (ultimate limit state) or a limiting deflection of the top deck beyond which
normal operation of the platform cannot be continued (serviceability limit state).

1.5 ROLE OF INSPECTION AND MAINTENANCE PROGRAMS

The inspection of the platform is carried out during its construction and subsequently
at periodic intervals during service. The purpose of inspection during construction is
to ensure that the platform is fabricated and installed in accordance with the designer's
1-7

plans and specifications. Most of the fabrication inspection relates to non-destructive


testing (NDT) of welds, alignments and nominal dimensions. Since it is almost
impossible to produce defect free welds fabrication specifications allow for some
"tolerances" on weld defects, dimensions and initial imperfections. For practical
reasons not all welds are inspected and varying levels of NDT are used for different
welds.

In-service inspection, on the other hand, is conducted to identify structural damage or


deterioration that may require remedial measures to ensure safety of the structure and
personnel. This type of inspection could involve a check on the adequacy of the
cathodic protection system, corrosion damage to structurals and production risers,
extent of marine growth, sea bed scour, damage to primary members and barge
bumpers near the water-line due to wave slamming and boat impacts, damage to
tubular joints due to overloading or fatigue crack growth. The inspection methods
involve visual inspection, NDT using magnetic particle inspection or eddy current,
ultrasonic thickness monitor or continuous monitoring using accelerometers, acoustic
emission and potential drop testing. Routine maintenance involves replacement of
missing sacrificial anodes, cleaning of marine growth, removal of debris etc., and
repair of corrosion damage or cracks are carried out when the damage is "significant".

The in-service inspection and maintenance programmes are generally very expensive
and the key issues involved are: which parts of the structure to inspect, what type and
level of inspection to use, how frequently to inspect and if damage is detected,
whether and when to repair and what type of repair (wet welding/hyperbaric welding
etc.) method to be used? Several issues have to be considered before arriving at a
decision and economics must play a role without compromising safety unduly.

From the foregoing discussion an inter-relation between design, inspection and


redundancy is immediately evident. Explicit allowances should be made during design
for unavoidable fabrication imperfections, welding defects and corrosion damage. The
stipulated extent, frequency, dependability of methods used and accessibility of
important structural components for inspection influence their initial design. On the
1-8
other hand, availability of redundancy. sound design, supplemented by good quality
control during construction reduce the need for and extent of in-service inspection.

1.6 CURRENT DETERMINISTIC DESIGN PRACTICE

The design of fixed offshore platforms is a complicated process involving interplay


of several factors. Separate design/analysis is carried out for fabrication,
transportation, launching, installation and in-place conditions. The final design is
influenced by such factors as material selection, welding and fabrication, inspection,
previous experience of the field and cost. It is not intended here to discuss all aspects
of offshore structure design. Only those aspects of the design which are relevant to
the subject of this thesis are summarized in the following and a detailed review of
individual aspects is presented in relevant chapters. In particular the design of a jacket
structure in the in-place condition against the action of static loads and environmental
forces due to wind, waves and current, is discussed; whereas the design of the deck
structures, the pile foundation and resistance of the platform against earthquake forces
is not considered.

A number of codes and standards have been developed which offer guidelines on the
design and construction of fixed offshore structures. The most widely used across the
world are those published by the American Petroleum Institute (1987), the U.K.
Department of Energy (1990) and the Norwegian Petroleum Directorate (1989). These
are supplemented by several other codes which are not developed for offshore
construction but are used for specific aspects of the platform design. In addition,
classification agencies such as Lloyd's Register of Shipping, Det-norske Veritas and
the American Bureau of Shipping publish their own "minimum" requirements for the
certification of fixed offshore structures. Together they represent the current "industry
practice" for the design of offshore structures.

Most of the present design codes (except NPD and DnV) use the so-called "Working
Stress Design" (WSD) concept In this approach "nominal" values of the working
1-9
loads are applied to the structure to determine stresses. The safety of the member is
ensured by providing adequate "safety margins" between calculated member stresses
and limiting (yield) strength of the member computed using nominal material
properties. The nominal values of the loads, the material properties and the single
"safety factor" on member strength are specified by the code for different design
conditions and failure modes.

The extreme environmental condition, to be used in combination with the specified


values of dead and live loads from the deck for the design of the "intact" structure,
is based on the wave height with a return period of 100-years and "associated" current
and wind, API (1987). Statistical data on wind, waves and current at the site is
collected, typically over a period of 3-5 years prior to the installation of the structure.
Statistical methods are used to extrapolate and determine the 100-year return wave
height. Usually sufficient data are not available to describe the joint distribution of
wave height, wave period, wind speed, current speed and direction.

Hydrodynamic forces on various members are calculated using Morison's equation


with an appropriate wave theory for particle kinematics by positioning the wave crest
in such a way as to maximise the total overturning moment or horizontal base shear
on the structure. The static response of the structure to applied loads is calculated
using linear elastic structural analysis, in most cases considering the non-linear
response of the pile foundation. Elastic flexibility of tubular joints is usually included
in the global analysis although not stipulated in the design codes. Global dynamic
analysis of the platform is carried out only for deep water structures with natural
periods longer than 3 seconds. The design sea-state considered should be such that
it "could" produce the design wave height considered for static response. A dynamic
amplification factor (DAF) is worked out which is a ratio of the global maximum
dynamic response to global maximum static response. In calculating member stresses
due to combined effect, the dynamic response is assumed to be in phase and in the
same direction as the static response and member stresses from static response are
simply amplified by the DAF to obtain combined stresses.
1-10
The nominal strength of bracing members is calculated from code specified design
formulae. These formulae are usually derived from a lower bound fit to set of test
results. Tests on tubulars cover a range of typical geometries and material properties
used in offshore construction and generally considerable scatter in results is observed.
Separate strength formulae, for example by the API (1987), have been specified for
tubular members loaded by axial tension and moment loads. For tubulars loaded under
axial compression strength checks are carried out for both overall "column buckling"
and "local buckling" of thin cylinders. A separate check is carried out for hoop
compression under hydrostatic load of sea water. The effect of initial imperfections
and residual stresses is implicitly included in the above equations. Interaction
formulae, again derived from test results, are specified for tubulars under combined
loading. The codes do not give any information to describe the post-limit behaviour
of tubulars.

The strength of tubular joints against axial tension, compression and moment loads is
also derived from test results. The test results at present cover only simple tubular
joints. Again, large scatter in test results is observed. Some of the recent tests on
tubular frames, see Lalani and Shuttleworth (1990), indicate that the behaviour of
tubular joints in a frame can differ appreciably from those of the isolated tubular
joints. Thus a large uncertainty exists about the strength of tubular joints in the real
structure.

The fatigue design of tubular joints is usually based on the so-called "S-N" approach.
Curves describing the relation between a constant amplitude stress range (S) and the
number of cycles to failure (N) at this stress range have been derived from extensive
fatigue tests on tubular joints with various geometries under both air and sea water
environment. The fatigue loading on offshore structures is usually determined from
a spectral analysis of the structure under a discrete number of sea-states which are
supposed to represent the long-term wave climate at the location of the structure.
Because of the various uncertainties involved in the fatigue analysis procedure and
scatter in S-N test data it is usual to design tubular joints for minimum computed
fatigue lives of 2 4 times the expected life of the platform.
1-11
Most offshore design codes do not specify explicit checks for the fracture resistance
of tubular joints. Fracture control is ensured only through specification of "tolerance"
levels on weld defects found during fabrication, the use of material with a high level
of notch toughness for joint cans and proper welding procedures. Fracture toughness
levels are specified in terms of the Charpy-V notch energy for various grades of steel
and service temperatures. Specimens are also extracted from welds and heat affected
zones of fabricated joints to check that they meet the specified minimum toughness
levels. Explicit fracture assessments are undertaken only when defects larger than
specified minimum sizes are found during fabrication inspection.

The design of offshore structures as stipulated by most present codes involves safety
checks of individual members and joints only. Explicit evaluation of the redundancy
and ultimate capacity of the entire structure is not made. The NPD code (1989),
however, stipulates that the structure should have adequate resistance against abnormal
load effects such that the structure experiences only "local" damage from such effects
and that the structure in the (locai) damaged condition should be able to withstand
design environmental loads without extensive failure. Neglect of system behaviour
results in the use of equal safety factors for all members regardless of how "critical"
the member is for the integrity of the structure.

The nature and extent of inspection during fabrication as specified by codes is mostly
derived from past experience and the designer is not expected to carry out any
engineering analysis to determine this. For in-service inspection and maintenance the
procedure is usually dictated by the requirements of the certification and insurance
agencies. Considerable variation is observed between different certification agencies
and different fields, strongly depending on past experience. In general, inspections are
classified into different levels and parts of the structure and frequency for each level
of inspection is specified. Rational procedures have yet to be developed to quantify
properly the results of inspection and to use them in the planning of subsequent
inspections.
1-12
1.7 NEED FOR PROBABILITY-BASED ASSESSMENT METHODS

In the design of offshore structures considerable sources of uncertainties are involved


in the prediction of the extreme environmental conditions to which the platform will
be subjected during its lifetime. For a given sea condition uncertainties exist in the
calculation of the hydrodynamic loading on the members owing to the uncertain extent
of marine growth, surface roughness, the uncertainties in the values of drag and inertia
coefficients required in force calculations and the simplifications made in the
derivation of water particle kinematics using a wave theory. The response to gravity
and environmental loading is affected by the uncertain soil properties, structural and
hydrodynamic damping and the finite element idealization of the structure. Thus the
predicted forces and stresses in the various members and joints of the structure are
subject to considerable uncertainty. The predicted strength of the components is also
subject to, uncertainty because of the inherent randomness in material properties such
as yield strength, fracture toughness, fatigue strength and the uncertain magnitude of
detrimental factors such as welding defects, residual stresses, dimensional and
alignment imperfects, etc. The use of simplified models for the capacity of tubular
members and joints add to the uncertainty in the predicted strength values.

In the current practice of offshore structure design based on the working stress
approach a set of "nominal" loads and material properties are specified by the code
0
for each mode of component failure associated with a "factor safety" on the member
strength. Thus load and resistance are taken as single valued (and deterministic)
quantities without regard to their variability. The specified safety factors are usually
derived from judgement and experience and not by any quantitative assessment of the
uncertainties involved and the resulting probability of failure. It can be shown that
when designing to a safety factor of as high as 2.5 the actual failure probability of a
component can vary from a low value to an intolerably high one, depending on the
nature of the variability of the design quantities.

Since most loads and material properties are random in nature there is a need for the
use of pmbabihstic concepts into structural design. It is now widely recognised that
1-13

in the presence of uncertainty absolute safety is unattainable and some risk of


unacceptable structural performance must be tolerated. The main object of structural
design is therefore to ensure, at an acceptable level of probability, that a structure will
not become unfit for its intended purpose at any time during its specified design life.
The modem methods of reliability based design help to achieve this objective.
Structural reliability theory provides a framework for the rational treatment of
uncertainties in structural design. In these methods the safety of the structure is
quantified in terms of the 'probability of survival' or 'reliability' which is now
generally accepted as a more rational measure of structural safety.
In recognition of the need for proper quantification of uncertainties involved in the
design of offshore structures the American Petroleum Institute has spent considerable
effort in the past few years in developing a new "LRFD" design code, API (1989).
This document has recently been released for review by the industry. This code uses
a "limit state design" concept and the load and resistance partial safety factors for
component design have been derived from a reliability analysis quantifying the various
sources of uncertainties. The safety factors have been calibrated to give, in general,
the same level of safety as implied by the present working stress design code for the
Gulf of Mexico wave climate. A similar effort is currently under way by the U.K.
Department of Energy to develop a LRFD code format for the North Sea environment.

The new LRFD codes offer considerable simplicity in routine design and provide more
consistent level of reliability among various members and under varying load
conditions. However, it should be recognised that they have been calibrated for
particular types of structures and particular wave environments and as such should not
be used for different types of structures such as compliant structures or
semi-submersibles and in different areas. Also they cannot reasonably be used in the
safety assessment of existing (possibly degraded) structures. In such non-routine
situations and in frontier areas, a direct use of structural reliability methods becomes
attractive. With the development of modem fast computers and considering the high
cost and high risk of offshore structures, the direct use of reliability methods in routine
design may become possible. Efforts should be spent to develop and improve existing
1-14

structural reliability methods to make them practicable for the reliability assessment
of offshore structures.

1.6 OBJECTIVES OF THE THESIS

From the foregoing discussion it is clear that jacket-type fixed offshore structures
exhibit complex behaviour in their response to environmental loading. The collapse
of the (entire) structure is influenced by the capacity and post-limit behaviour of
individual members and joints, topology and degree of redundancy. The design of
these structures involves considerable sources of uncertainty in the prediction of
loading and strength of individual members and the structure. The use of modern
methods of structural reliability analysis is, therefore, appropriate in ensuring safety
of these structures.

The conventional methods of structural design have concentrated on strength checking


of individual members and the safety of the entire structure is not assessed
quantitatively. Depending on the degree of redundancy present in the structure, this
may lead to either over-conservative design or catastrophic failure due to the failure
of one or two elements under accidental loads. For rational assessment of safety it
is important to design for a system limit-state giving due account for the redundancy
and ultimate strength of the structure. This requires the use of system reliability
methods and in general identification of all the significant modes in which a structure
can fail. An efficient method for the non-linear progressive collapse analysis of
structures is therefore vital for the successful use of system reliability methods.

Considerable research effort has been made in recent years on the application of
system reliability methods to fixed offshore structures. These efforts have mainly
concentrated on the reliability analysis of the structure under extreme wave loading.
Although considerable progress has been made in this direction the studies reported
so far are inherently limited to idealized member behaviour (e g. elastic-perfectly
plastic etc.) and the treatment of wave loading is also simplified in most cases. The
1-15

progressive collapse of the structure is commonly determined by successive elastic


analyses with member replacement which makes these methods less accurate,
computationally expensive and limits the probabilistic analysis to consider only one
failure mode for each structural member. The behaviour of tubular joints is therefore
usually not considered and fatigue deterioration of the structure is not accounted for.
These effects are very important and should be included in any realistic assessment
of reliability of fixed offshore structures. On the other hand, a few studies are also
available on the analysis of a sequence of fatigue failures but these completely ignore
the possibility of collapse of the structure under an extreme wave.

The first objective of this thesis is, therefore, to develop a more efficient and practical
procedure for the system reliability analysis of offshore structures for the extreme
wave condition which can treat more realistic member behaviour with softening/
hardening effects. This is achieved by using a virtual distortion method of non-linear
structural analysis with the joint beta-point formulation for reliability analysis.

The next objective is to extend the above approach to account for tubular joint
behaviour and the effect of fatigue deterioration on the reliability of the structure.
This approach, however, is essentially "time-independent" in the sense that all the
elemental failures, leading to the collapse of the deteriorated structure, are assumed
to occur under a single extreme wave.

In reality, fatigue is a time-dependent phenomenon because of which various elemental


failures can occur at different points in time. The third objective is to develop an
alternative formulation of time-dependent system reliability analysis for the treatment
of a sequence of fatigue/fracture failures in time.

The final objective of the thesis is to propose an unified approach in which


quasi-static elemental failures under an extreme wave and time-dependent
fatigue/fracture failures can be combined in a single analysis to estimate the life-time
reliability of the structure.
1-16
In order to achieve the above objectives realistic mechanical models are required to
predict fatigue crack growth in tubular joints using a fracture mechanics approach, the
collapse strength of damaged joints and the ability to assess the fracture criticality of
a joint with a given crack. A detailed review of literature reveals that prediction of
tubular joint behaviour is a topic of intense on-going research and a fully satisfactory
and consistent model taking account of all these factors is not yet available.
Considerable effort is therefore required to develop consistent and simple models of
tubular joint behaviour which can treat most of the design random variables in an
explicit way so as to be suitable for use in reliability analysis. The available methods
of spectral analysis to determine fatigue loading on offshore structures are well
developed, although some improvements in the treatment of broad-banded stress
spectra are considered necessary to increase efficiency and accuracy. The first step
in a system reliability analysis under fatigue is to develop a model for reliability
analysis of individual tubular joints considering the various uncertainties in the
prediction of fatigue loading due to waves and the failure behaviour of joints.

In view of the above, the main tasks set out for this thesis are as follows:

* To describe, probabilistically, the long-term and extreme environmental


conditions that may occur during the life of the structure and obtain the
response of the structure to these conditions. In particular to develop a
probability density function model for stress cycles of broad-banded response
spectra and to compute cumulative fatigue loading.

* To evaluate existing models and develop new models to predict fatigue crack
growth in tubular joints, collapse strength of damaged joints and to assess
fracture criticality of joints with cracks.

* To develop methods to quantify the various sources of uncertainty in the


loading and capacity of tubular joints and to carry out reliability analysis of a
tubular joint for the limit states of fatigue failure, fatigue-fracture and plastic
collapse.
1-17
* To develop a more practical method for the system reliability analysis of the
structure under an extreme wave accounting for strain-dependent post-limit
behaviour of tubular members. Associated developments within the CEC
BRITE Project No. P1270 has been made use of and a number of researchers
have contributed in accomplishing this task. In particular, a response surface
model for environmental loading and a virtual distortion method of non-linear
structural analysis have been used.

* To extend the above approach to account for tubular joint behaviour and the
effect of fatigue deterioration on the reliability of the structure.

* To develop an alternative formulation for the system reliability analysis of the


structure in which tubular joint failures due to fatigue and fracture occur at
different (random) points in time.

* Finally, to develop an unified approach to combine fatigue/fracture failure


sequences and quasi-static elemental failures under an extreme wave in a
single analysis to estimate life-time reliability of the structure.

The scope of the thesis is restricted to the study of the jacket-structure of a fixed
offshore platform subjected to gravity, wind, wave and current loading. For a
complete assessment of the reliability of the platform the deck structure and pile
foundation should be included and other forms of loading such as earthquake forces
should be considered. In addition, other design conditions such as fabrication,
transportation, launching and installation should be separately checked for. The
emphasis of this study is on fatigue behaviour of tubular joints and their consequent
effect on the reliability of the system. The behaviour of tubular members is not
discussed in detail. In the discussion to follow it is implicitly assumed that the
structure considered is a shallow water jacket platform for which dynamic (inertial)
effects are not significant.
1-18
1.9 ORGANISATION OF THE THESIS

The thesis is organised into 8 chapters, the first being the present introduction. The
last chapter presents some of the important conclusions of the thesis and topics on
which further research effort is needed. The content of the remaining chapters is
outlined in the following.

The basic theory of reliability analysis and the various computational tools that are
available are reviewed in Chapter 2. General expressions for the 'probability of
failure' or 'reliability' of individual components and that of series systems with
parallel sub-systems are given and the use of first order (FORM) and second order
(SORM) solution methods is examined for their limitations, efficiency and accuracy.
A number of sensitivity and importance measures reported in the literature are
examined for use in judicious selection of important basic variables and components
to reduce the computational effort involved in the reliability analysis of large systems
such as offshore jacket structures.

In order to carry out reliability analysis it is first necessary to develop appropriate


probabilistic models for the basic load and resistance variables. Chapter 3 of the
thesis is devoted to modelling of the loading and response of offshore structures.
Probabilistic modelling of long-term sea surface fluctuations as required for a fatigue
analysis is discussed. Stress response of individual members and joints is obtained
using standard methods of spectral analysis of offshore structures. A new probability
density function is derived for modelling stress cycles in a broad-banded stress process
and computation of cumulative fatigue loading. Suitable models are also presented
for the extreme environmental condition and extreme response that may occur during
the service life of the structure.

Mechanical models for the capacity of tubular joints arepresented in Chapter 4 of the
thesis. A fatigue crack propagation model is developed, based on a fracture
mechanics approach, to quantify the extent of fatigue damage in tubular joints. This
model explicitly accounts for the important influences of fatigue threshold, residual
1-19
stresses, weld geometry, weld toe plastic zone and crack coalescence. Under an
overload situation a tubular joint with a fatigue crack could fail either by plasticity or
fracture instability of the crack. The available empirical models for collapse loads of
intact joints are examined and a simplified method for estimation of the strength of
damaged joints is suggested. A fracture assessment model for tubular joints is
proposed based on the CEGB-R6 design approach.

Chapter 5 presents the reliability analysis of a tubular joint of an offshore structure


considered in isolation from the rest of the structure. Treatment of various sources
of uncertainties in the modelling of environmental loading, response prediction,
inherent randomness in material properties and initial weld defects and uncertainties
in strength prediction models is discussed in detail. Reliability analysis is carried out
for the limit states of fatigue failure, fatigue fracture and plastic collapse. First
order reliability analysis methods are used and the sensitivity of the reliability index
to basic random variables is examined over a range of fatigue exposures of the joint.
Important basic variables and dominant failure modes of the joint are identified for
subsequent use in system reliability analysis.

Reliability of an offshore structure under the life-time extreme wave condition is


considered in Chapter 6. The influence of post-limit behaviour of primary members
and joints on the ultimate strength of the structure is examined and a number of
deterministic methods for the progressive collapse analysis of the structure are
reviewed. Limitations of the conventional first order reliability analysis method used
in conjunction with a "member replacement" approach and plastic mechanism based
methods are examined. A new reliability analysis method based on the "joint-3 point"
concept in conjunction with a virtual distortion method of progressive collapse
analysis is proposed. A modified algorithm for non-linear structural analysis for a
prescribed failure sequence based on the virtual distortion method is presented. To
identify stochastically dominant failure modes a modification of the
"branch-and-bound" procedure is proposed which is suitable for practical analysis of
large structures. The above procedure is extended to incorporate fatigue effects in the
form of deterioration in tubular joint capacities under plastic collapse and fracture
1-20

failure modes. An example analysis of a shallow-water jacket structure is presented


to demonstrate the proposed methodology.

In the above approach for treatment of fatigue effects, however, all the elemental
failures, leading to collapse of the deteriorated structure, are assumed to occur under
a single extreme wave. Since fatigue is a continuous process it is likely that one or
more joints could fail under different storm conditions. An alternative time-dependent
system reliability analysis method is developed in Chapter 7 of the thesis. A general
formulation is first given for a time-dependent failure path in which componential
failures are considered to occur at different random points in time under multiple wave
peaks. Application of this to system reliability analysis of a jacket structure
considering only fatigue failures or fatigue followed by fracture failures is presented.
Finally, a unified approach for combining both time-dependent and extreme load
failures is proposed. This method allows the development of a single failure-tree for
the structure which includes: (i) pure quasi-static extreme load failure sequences in
which all elements fail under a single large (life-time extreme) wave, (ii) pure fatigue!
fracture failures in which brittle elemental failures occur at various points in time
during the service-life of the structure and (iii) mixed failure modes in which initial
failures occur due to fatigue/fracture at random points in time and the significantly
weakened structure then collapses under a single large wave.

The work described in this thesis has been carried out under the over all frame-work
of the CEC BRITE Project No. P1270 and the proposed models have been
implemented into an operational computer software called RASOS (Reliability
Analysis System for Offshore Structures) in association with other researchers. An
outline of the RASOS software is given in Appendix-A of the thesis. The contribution
of the author has been mainly in modules RASOS_H, RASOS_F and RASOS_R. The
RASOS software has been used for the reliability analysis of a shallow-water jacket
structure the results of which are presented in parts in chapters 5, 6 and 7.
Chapter 2
A CRITICAL REVIEW OF COMPUTATIONAL METHODS FOR
STRUCTURAL SYSTEM RELIABILITY ANALYSIS

2.1 GENERAL

As emphasized in Chapter 1, a considerable number of sources of uncertainty are


involved in predicting the behaviour of jacket-type offshore structures and to ensure
sufficient safety it is important to use probability based methods in the design and
assessment of these structures. Structural reliability theory, which has been a field of
intense research and development over the last two decades, holds considerable
promise in this respect. The purpose of this chapter is to introduce some of the
fundamental concepts of the theory and thus set out a logical framework for the
discussion to follow in subsequent chapters.

It is not possible here to discuss in detail all aspects of the system reliability theory.
An attempt is, however, made for a self-contained presentation of the various
computational methods that are used in later chapters in the context of reliability
analysis of offshore jacket structures. More information about these and other
available computational tools can be seen in standard texts on the subject, for example
Thoft-Christensen and Baker (1982), Madsen, Krenk and Lind (1986) and
Thoft-Christensen and Murotsu (1986).

The term 'reliability' is first introduced as a probabilistic measure of structural safety


and the steps involved in the reliability analysis of a structural system are outlined.
Reliability analysis of individual structural elements is discussed with a presentation
of the so-called FORM and SORM methods. Modelling of system failure events in
terms of limit-states is presented for series systems, parallel systems and general
systems. Some of the computational tools available for the reliability evaluation of
these systems are critically examined with regard to their efficiency and accuracy.
Finally a number of sensitivity measures are introduced and their usefulness in
practical applications is highlighted.
2-2
2.2 PROBABILISTIC BASIS OF STRUCTURAL SAFETY

The purpose of structural design is to produce a safe and durable structure at optimum
cost. The question of safety arises because of the uncertainty about the magnitude of
loads the structure will be subjected to during its lifetime and the ability of the
structure in resisting these imposed loads, in the conventional design practice, either
using a 'Working Stress' approach or a 'Limit State Design' approach, safety of a
structural component is ensured by a choice of "expected" or "characteristic" values
for the design load and resistance variables combined with one or more "safety
factors". Safety of the overall structure is not usually checked explicitly and the
structure is deemed safe if all its components are shown to be safe. Since the
underlying uncertainties in the design process are not quantified the safety factors do
not provide a rational measure of structural safety.

In a probabilistic approach to design, it is well recognised that in the presence of


uncertainty absolute safety cannot be achieved. The purpose of design is then to
ensure that the probability of a structure becoming unfit for use within its intended
lifetime is "acceptably small". In structural reliability theory 'probability of survival'
or 'reliability' is taken as a quantitative measure of structural safety. The uncertainties
involved in analyzing the performance of a structure are accounted for in estimating
the probability of failure or, its complement, the reliability of the structure, which is
now widely accepted as a more rational measure of structural safety. The partial
safety factors used in some of the recent European and Canadian codes for the design
of buildings and bridges have been derived based on a reliability approach.

Although the probability of failure as defined above can be interpreted in a relative


frequency sense, as the percentage of a notionally infinite set of nominally identical
components, the reliability of a structure can rarely be determined by observation of
the population. This is because each structure is unique and the population then
becomes abstract. The reliability of a structure is, therefore, calculated from the
probabilities of failure of individual structural elements. These, in turn, are evaluated
from system parameters, global loads on the structure, component stresses and strength
2-3
characteristics of components as a function of random material properties. In this
sense structural reliability theory is quite distinct from the classical reliability theory
used for mechanical or electrical components or systems.

An important source of uncertainty which is a cause of most structural failures is


'human error'. The causes of these errors are psychological and social factors which
are complex and not yet well understood. Human errors, to some extent, can be
controlled by better working conditions, rigorous quality control and training etc.
Uncertainties due to human error are not generally accounted for in a reliability
analysis. Thus the result of a reliability analysis should be considered as a 'theoretical
value of the reliability of an "error-free" structure computed from a reliability model'
and is different from the observable failure frequency for homogeneous, mass
produced items. It should also be noted that reliability is not a property of the
structure but a reflection of the analyst's present state of knowledge about the
structure and its value changes as more information becomes available. However, this
does not reduce the merit of reliability as a rational measure of safety. Using a
consistent reliability model the theoretical value of reliability can be used by the
designer as a decision tool to compare reliabilities of two components or in choosing
between two structural alternatives when considerable source of uncertainties are
involved. Reliability methods become immensely important in the safety assessment
of existing structures and when novel structures or new frontiers are involved where
little previous experience is available.

Complex structural systems such as offshore jacket structures consist of a large


number of structural elements, each of which can fail in a number of different ways.
Because of the considerable degree of redundancy there is usually a large number of
possible combinations of element failures that can result in the failure of such
structures. In these cases it is not a simple matter to evaluate the overall reliability
of a structure and idealizations are necessary in representing the system behaviour of
such structures. Depending on the failure behaviour of structural elements a number
of 'limit-states' are identified for individual elements and reliability analysis is first
carried out for each of these limit-states. These element limit-states are sometimes
2-4

referred to as 'failure elements'. In a redundant structure, a sequence of element


failures leading to the collapse of the structure constitute a 'failure mode' of the
structure. Because of random loads and component strengths, potentially a very large
number of failure modes are possible for a structure. The system probability of failure
is then defined as the probability that the structure can fail in any one of these modes.
Some of the general principles involved in these idealizations and the necessary
computational tools for reliability analysis of individual elements and of systems is
presented in various sections in this chapter.

2.3 TREATMENT OF UNCERTAINTIES IN A RELIABILITY ANALYSIS

As mentioned previously, uncertainties in structural design arise from the random


nature of material properties and applied loads. Uncertainties can also arise because
of the use of imperfect and simplified methods for the prediction of imposed stresses
and component strengths. In general, uncertainties can be considered to be of three
types, namely: physical uncertainty, statistical uncertainty and model uncertainty,
Thoft-Christensen and Baker (1982).

Physical uncertainty arises from the inherent randomness of physical quantities, such
as loads, material properties and dimensions. This variability can be described in
terms of probability distributions or stochastic processes developed from observed
statistical data.

Statistical uncertainty is due to limited sample sizes in the estimation of distribution


parameters of a design quantity. Sample sizes are often limited by practical and
economic consideration. Thus statistical uncertainty arises solely as a result of lack
of information about a variable. Thus for a given set of data the parameters of the
distribution describing a variable could themselves be considered as random variables.
The problem is more pronounced when the variable exhibits non-stationarity, since in
this case statistical data collected at one or two time instants may not correctly
represent the long term variation of the variable.
2-5
Model uncertainty is associated with the mathematical models used, such as for stress
analysis and component strength estimation. These mathematical models are usually
deterministic but may also be probabilistic as in the calculation of stress response
spectra of offshore structures subjected to wave loading. A number of idealizations
and approximations are usually made in these mathematical models. In some cases
prediction models are developed from empirical fits to experimental data, for example
stress concentration factor formulae for tubular joints. The use of such simplified
models gives rise to model uncertainties.

Depending upon the problem being studied uncertainties can arise from a large
number of sources all of which may not be amenable for systematic analysis and
estimation. In a reliability analysis important sources of uncertainty are quantified
through a finite number of quantities called 'basic random variables'. They are basic
in the sense that they are the most fundamental quantities normally recognised and
used in routine design calculations. Thus they may include material properties such
as yield strength and fracture toughness, loads due to gravity, wind and waves and
model uncertainty variables. One simple guideline that may be used in choosing a
basic variable is that enough statistical data can be collected about it. Depending on
the amount of data available, the uncertainties in design can be described in terms of
the sample mean and variance for each basic variable or more rigorously in terms of
the joint probability density function of all basic variables and its parameters.

2.4 ELEMENT RELIABILiTY ANALYSIS

In order to assess the safety of a structural element, a limit-state or a set of limit-states


are identified which determine the performance of the component. These limit-states
could be either ultimate limit-states or serviceability limit-states. For each limit-state
a reliability analysis is carried out to determine the probability that this limit-state will
not be reached. The reliability of the structural element is then defined as the
probability that none of the limit-states will be reached within the specified service life
of the structure. Reliability formulation for a single limit-state is presented below.
2-6

2.4.1 Problem Formulation

The first step in a reliability analysis is to identify a set of basic random variables
which influence the failure mode or the limit-state under consideration. Let in the
following X = (X1 , X2 ......, X) represent a vector of 'n' basic variables.

Next, a 'limit-state function' or a 'safety margin equation' is formulated in terms of


the n basic variables, Thoft-Christensen and Baker (1982),


M - g(X) - (2.1)

in such a way that g() satisfies the following:

whenEo3
(2.2)
whenEo1

setting

g(i) - g(x1,x2,...,x) - 0 (2.3)

defines an (n-i) dimensional surface in the space of n basic variables. This surface
is called a 'failure surface' or a 'limit-state surface' and divides the basic variable
space into a 'safe region' u and an 'unsafe region' (Of. Note that the limit-state
function g() is entirely deterministic and any of the existing deterministic models for
strength prediction could be used. The random variable M is called a 'safety margin'.

The reliability or the probability that the limit-state will not be reached is then
expressed as

R - l-p1 - 1-P[M^O] - 1-fr()di (2.4)

where f() is the joint probability density function of the n basic variables and p
denotes probability of failure. The n-dimensional integral is defined over the failure
region (Of.
2-7
In practical applications the reliability cannot be evaluated in the "exact" manner as
given by Eq.2.4. This is firstly, because enough statistical data is usually not available
to develop the n-dimensional joint density function of the basic variables. Secondly,
even when the joint density function is available, analytical or numerical integration
is possible only for a few simple cases. For practical applications a number of
simplified solution methods have been developed. These methods, discussed next,
make varying degree of simplifications in the probabilistic representation of basic
variables or in evaluating the reliability.

2.4.2 Second Moment Reliability Indices -

In order to characterize the uncertainty in a basic variable in terms of its probability


density function a large amount of statistical data is necessary. In practical situations,
usually only limited data are available with which it is possible to estimate only the
mean value (first moment) and variance (second moment) with confidence. In such
cases reliability methods which employ up to second moment information to
characterize uncertainty in a basic variable are required. Since, in this case, the
probability distributions of the basic variable are not known it is not possible to
evaluate the probability of failure of a structural element. A number of indices were
therefore developed, which provide a relative measure of reliability. These indices
which are based on first and second moment representation of basic variables are
called 'second moment reliability indices' and the reliability methods which only
compute a reliability index and do not attempt to evaluate the "exact" probability of
failure are called 'Level 2 methods' or 'second moment reliability methods',
Thoft-Christensen and Baker (1982) and Madsen, Krenk and Lind (1986).

The earliest definition of a reliability index was given by Cornell (1969). In this
approach a limit-state function or safety margin equation is first formulated in terms
of the basic variables as discussed before. Then the reliability index is defined as

(2.5)
C PM/M

2-8
where, IM is the expected value and GM is the standard deviation of the safety
margin M. f3 is called the 'Cornell reliability index'. The reliability index as defined
above simply gives the distance from the mean value point to the failure surface in
units of the uncertainty scale parameter and is thus believed to provide a good
identification of safety. Thus for a linear safety margin, as given below, M' 0M and
hence f3 can be calculated as

M— a 0 ^ a 1X1 ^ .... + a,X (2.6)

a0 + ai p. x + .... +
- - -
fl lt
G M 22 (2.7)
[a 1 ax ^ .... + aG + E E
i-I •1
1

where Pu = Co v [Xj,X]/a,(Yx is the correlation coefficient between the variables X


and Xi . However, when the limit-state equation is non-linear in terms of the basic
variables the mean value J1M and standard deviation GM cannot be calculated from the
information of first and second moment of X alone. This problem can be overcome
if the limit-state function can be linearized suitably. This makes a choice of
linearization procedure necessary. Thus for a general non-linear failure function,
defined below, using a Taylor series expansion around a point x = (x4, x2,...x) a
linearized safety margin equation can be obtained. For the linearized safety margin
equation, GMS and hence the reliability index can be calculated as before. The
reliability index calculated in this way is sometimes called a 'first-order,
second-moment reliability index', I3fm• Thus for

M - g(X)

(2.8)
'I
(2.9)
M, - g (x) +
E
2-9

g(x) + ._(j.L1-x1)
R
1'fo:m - (2.10)
[a a /2

EE
i-i j-1
aCOV[X,X]J

If the linearization is canied out about the mean value point, then the reliability index
calculated as above is called as a 'mean value, first-order, second-moment reliability
IndeX, J3mvfosm

One of the principal drawbacks of the reliability indices defined by Eqs. 2.7 and 2.10
is known as 'lack of failure function invariance'. This means that, for a non-linear
limit-state function of a physical problem, different but mathematically equivalent
formulations of the limit-state function lead to different values of the reliability index.
This was first explained by Ditlevsen (1973). This was because the reliability indices
were related to the limit-state function and not the failure surface. Since all
mathematically equivalent limit-state functions lead to the same failure surface, for a
given physical problem, the reliability index should more appropriately be related to
the failure surface. This realization led to the 'Hasofer and Lind reliability index'
discussed next.

In order to relate the reliability index to the failure surface, Hasofer and Lind (1974)
proposed, first, a linear transformation of the set of basic variables X into a set of
standardized and uncorrelated variables Y. This transformation can be represented as

L: X - (XPX2,•••,Xa) (2.11)
' ' - ('I''2'"')

E [X ],Cj - Cov[X,X T] —. E[Y] - 0 and C - Cov[Y,YT] -

where C is a matrix of covariances between the variables, 0 is the null matrix and I
is the unit matrix. With this transformation the failure surface, g()=0 in the original
X-variable space is mapped on to a failure surface g(5)=0 in the Y-variable space.
Also, the mean value point in X-space is mapped into the origin of the transformed
space. It is, furthermore, to be noted that any orthogonal transformation of the Y-set,
2-10
results in a new set of standardized, uncorrelated variables. The distribution of the
Y-set is thus rotationally symmetric. More about how to obtain this transformation
is discussed in Section 2.5. The Hasofer and Lind reliability index is now defined as
'the shortest distance from the origin to the failure surface in the standardized
Y-coordinate system'. This definition can be formulated as

r 11/2
21 (2.12)
IHL - mm E 1]
L-

The point on the failure surface closest from the origin is usually called a 'design
point', . This design point can be mapped back to point i on the failure surface
in the original basic variable space by using the appropriate inverse transformation of
Eq.2. 11. In the original variable space the reliability index simply measures the
distance from the mean value point to the design point 1. Several algorithms are
available for locating the design point, see Thoft-Christensen and Baker (1982) or
Madsen, Krenk and Lind (1986).

It can easily be seen that for a limit-state function which is linear in Y-space, i.e. a
hyperplane failure surface, the value of IIL coincides with the value of 13c.
Furthermore, for a non-linear failure surface, the value of J3 will be the same as for
an approximating tangent hyperplane at the design point in the standardized Y-space.
Thus the problem of lack of failure function invariance can be avoided ii in Eq. 2.9
the linearization is carried out at the design point i. In this case the value of i3fm
will coincide with .

As mentioned earlier, the aim of defining these second moment reliability indices is
to be able to compare safety levels of different structures and order them based on
their reliability indices. The Hasofer and Lind reliability index generally provides a
good ordering of structures in most practical cases. However, when the limit-state
function g highly non-linear or the principal curvatures of the failure surface at the
design point are too large, the t 3 does not provide a satisfactory ordering. This is
because the definition of l 3 does not take account of the true shape of the failure
surface.
2-11

To overcome the above problem Ditlevsen (1979a) introduced a second moment


reliability index called the 'generalized reliability index'. The approach here, as
before, is to first transform the basic variables into a set of standardized and
uncorrelated variables Y. The new failure surface now divides the Y-space
into a safe region o and a failure region 0½. An arbitrary weight function t(5) is
now introduced in Y-space and a good measure of reliability is obtained by integrating
this weight function over the safe region (or alternatively a measure of probability of
failure is obtained by integrating this weight function over the failure region of). The
generalized reliability index is then defined as

I3 G - - -b'[j16)dw,] (2.13)

For reasons of simplicity and the property of rotational symmetry the weight function
chosen is the n-dimensional standard normal density function (pCy) with zero mean
and unit variance for which many solutions are available. The generalized reliability
index, since it takes account of the true shape of the safe region provides a good
ordering of safety levels of structures but is more difficult to evaluate. Some of the
methods for evaluating the n-dimensional standard normal integral in which the safe
region is approximated by a set of hyperplanes or a set of quadratic surfaces are
discussed in a later section on system reliability methods. It is to be noted that for
hyperplane failure surfaces coincides with I3 and for non-linear failure surfaces
with a single local minimum at which the principal curvatures are not too large, values
of 13u and 13 are close to each other.

In the definition of the generalized reliability index, the n-dimensional standard normal
density function is chosen only as a mathematical formalism and does not imply a
normal assumption for the distribution of basic variables. The rotational symmetry
property of the normal distribution provides a consistent ordering of safety levels of
different structures in terms of their generalized reliability indices. However, as
pointed out by Der Kiureghian and Liu (1986), this symmetry property is valid only
when the information on the basic variables is limited to first and second moments.
2-12

Generally, any additional information, such as higher moments, bounds or marginal


distributions, disturbs this rotational symmetry and hence the generalized reliability
index should be used strictly in a second moment context.

Alternatively, if a suitable joint probability distribution is assumed for the basic


variables such that this distribution is consistent with all the available information on
basic variables (including more than second moment information) a transformation to
the rotationally symmetric standard normal space is possible. Thus, when more than
second moment information is available the concept of generalized reliability index
can be used in conjunction with a formal distribution model for the basic variables and
a formal transformation of the basic variables to the standard normal space. Der
Kiureghian and Liu (1986) propose some distribution models and transformations
depending on the available information on basic variables. Reliability calculations in
these cases will be similar to those discussed in the next Section.

2.4.3 Evaluation of Failure Probability using FORM and SORM

The reliability indices discussed so far use information only about the first and second
moments of basic variables. Although they are useful in the ordering of safety levels
of different structures they do not give any indication of the actual probability of
failure or reliability as defined in Eq. 2.4. They also do not provide a mechanism for
incorporating any information beyond the second moments of variables. In practical
applications when sufficient data are available appropriate probability distributions
may be constructed. In some cases suitable probability distribution may be prescribed
for a basic variable based on physical or mathematical considerations, e.g., the normal
or lognormal distribution based on the central limit theorem, extreme-value
distributions based on asymptotic theorems. When information about probability
distribution of basic variables is available it should be possible to evaluate the
probability of failure or reliability of a component exactly. However, exact evaluation
of reliability, as in Eq. 2.4, by multi-dimensional integration of the safe region of a
generally non-linear limit surface is computationally very expensive For structural
2-13
components of high reliability it is generally felt sufficient to calculate the failure
probability within a factor of say, 2 to 10, Madsen et a! (1986). This section presents
methods for evaluating the probability of failure efficiently with reasonable accuracy.

When the failure surface is a hyperplane and the set of basic variables are jointly
normal it can be shown that the Cornell reliability index is related to the exact
probability of failure

(2.14)
pj - -

where c1( .) is the cumulative distribution function of a standard normal variable and
I1() is its inverse function. When the failure function is nonlinear p can be
obtained exactly by substituting the generalized reliability index for c in the
above equation. In this case the use of Hasofer and Lind reliability index f3 gives
a first-order approximation to p . since it approximates the actual nonlinear failure
surface by a tangent hyperplane at the design point in the standardized space.

The above results for the special case of normal variables suggest a simplified
approach for the evaluation of probability of failure in the general case of basic
variables with any arbitrary distributions. The approach is to first transform the basic
variables to a set of independent, standardized normal variables using an appropriate
transformation and locate the point on the failure surface, in the transformed space,
which is closest from the origin, i.e. the design point. A first-order approximation to
the probability of failure can be obtained by linearizing the failure surface at the
design point. This procedure is called a first order reliability method (FORM).
A better estimate of the probability of failure can be obtained by approximating the
failure surface by a quadratic surface at the design point. This procedure is called a
second order reliability method (SORM). Since FORM and SORM methods evaluate
the probability of failure using complete probability distribution information of basic
variables, these methods are categorized as 'Level 3 methods' by Madsen et a! (1986)
to distinguish them from second-moment Level 2 methods. However, since the
complete failure surface is not described and simple approximations to the failure
2-14

surface are used at one or more points in evaluation o probability of failure, these
methods are categorized as 'advanced Level 2' methods by Thoft-Christensen and
Baker (1982). In this thesis these methods are simply referred to as FORM and
SORM methods to avoid any confusion. These methods are reviewed next.

2.4.3.1 First Order Reliability Method

The general procedure of a first order reliability method may be described as follows.
Let X = (X1,X2,...X) be a set of n random basic variables. As in Section 2.4.1, for
a mode of failure, a limit-state function g(X) is defined in terms of the basic variables
in such a way that g(i)^ 0 for in failure set, g(i)= 0 for on the limit-surface and
g()> 0 for in safe set. Furthermore let M = g(X) define the safety margin for the
limit-state. Then the probability of failure is given by Eq. 2.4.

In a first order reliability method, see Madsen et al (1986), the set of basic variables
X is first transformed to a new set, U=(U1,U2,...,U) using a one-to-one transformation

T: X - - U - (U1,U2,...,U) (2.15)

such that the new set of variables are independent, standardized and normally
distributed. Some of the transformation methods are described in Section 2.5. The
equation for the limit-state surface in u-space becomes


g ) - g(T - '()) g(2) - 0 (2.16)

In the transformed space, a point on the failure surface closest from the origin, i1 is
determined by a minimization solution with one constraint such that
2-15

1s: nilnU; with g ) - 0 (2.17)

The point ii is called a 'design point' or a 'most likely failure point' as it represents
a point of highest probability density in u-space on the failure surface. At the design
point the unit normal vector or the direction cosines to the failure surface with respect
to each of the variables are

Vg(u )
a - - ________ (2.18)
I Vg(u)I

These direction cosines are often referred to as 'Sensitivity factors' in reliability


literature, see Thoft-Christensen and Baker (1982). More about their use as sensitivity
indices is discussed in Section 2.7.

Linearization of the failure surface at the design point gives an approximating tangent
hyperplane defined by

ag S
.__I;_;•(u—u ) - 0 (2.19)
r

and the shortest distance to this hyperplane from the origin is determined by

n
(2.20)
13 - _aT;*_Izu*2]
Li-i

Then the first order approximation to the failure probability is computed as


2-16


Pt- T(-l3) (2.21)

and (3 is usually termed as a 'first order reliability index'.

In terms of and (3, equations for the approximating hyperplane (Eq.2. 19) can be
re-written as


E au1 + - 0 or + 13 - 0 (2. 19a)

In principle any optimization algorithm can be used to search for the design point in
u-space. An algorithm with a fast rate of convergence is given below. In this
algorithm iterations are carried out in the u-space. However, in most practical
applications the limit-state function has to be formulated in the space of x-variables
as these represent design quantities in a deterministic model describing the response
of the structural component in the failure mode considered. Thus, at each iteration
step the u-variables have to be transformed to the x-space in order to determine the
safety margin. The algorithm can be stated as follows:

1) Select a trial vector ii°, preferably in the region of ii if previous knowledge


about the design point is available. In most cases the mean value point,
= 0, can be taken as a good starting value.

2) Transform the li-vector to the original space of i-variables using the


appropriate inverse transformation of Eq.2. 15 as discussed in Section 2.5.

3) Determine the safety margin g(T 1 (ii)) and the derivatives of g(il) with respect
to each of the basic variables. In most applications these derivatives have to
be evaluated numerically, incrementing each u-variable in turn and determining
the incremental safety margin after transforming the set of variables to x-space.
Evaluate the unit vector of direction cosines E using Eq. 2.18
2-17

4) Construct a sequence U1' U2,., 'sm'" by repeating steps (2) and (3) and at
each iteration (say mlh) determining better estimates of u-variables using, see
for example Gollwitzer et al (1988)


[ (2.22)
Um+ jg)]am
m

5) When the sequence converges to the design point 11' determine the reliability
index and probability of failure using Eqs.2.20 and 2.21.

Transformation of variables from u - x-space in step (3) may in some cases cause
considerable problems, especially if the inverse transformation is to be computed
numerically. An approximate linear backward substitution of 11m+1 into m+1 is
suggested by Hohenbichier and Rackwitz (1981). Alternatively, the entire algorithm
can be formulated in the space of x-variables. Such an algorithm is given by
Rackwitz and Fiessler (1978). This procedure works more efficiently for normal
correlated variables or non-normal independent variables.

A number of studies examining the accuracy of the first order reliability method have
been reported in literature, see for example Dolinski (1983). In most cases of
practical applications FORM is seen to give reasonably accurate estimate of the
probability of failure. The accuracy of the FORM method derives from the rotational
symmetry property of the standard normal space and because of the fact that the
probability density function of the variables decreases very quickly, namely as
exp(-r2/2), with the distance r from the origin. The area of integration giving
maximum contribution to the failure probability is therefore located in the region of
the design point and the failure surface is well approximated by a tangent hyperplane
around this point. However, when the principal curvatures of the failure surface at the
design point are large (i.e. highly curved failure surface) a tangent hyperplane
approximation is not satisfactory.
2-18

Failur. R.gion
(JI

Safe Region
CI•

Failure Surface
I.I' (u).O

TangenL Hyperplan.

FORM

U1

Fig. 2.1: illustration of probability of failure calculation using FORM.

From Fig. 2.1 it can be seen that, for convex failure surfaces (with respect to the
origin) FORM overpredicts the probability of failure while for concave failure surfaces
pf is underpredicted. Unfortunately FORM procedure does not provide a means for
estimating the level of enor involved.

The algorithm for FORM, presented above, also suffers from a few drawbacks. First
of all, the algorithm only locates a stationary point on the failure surface. Thus,
sometimes the algorithm could converge to a maximum point or when there are a
number of local minima on the failure surface, it could converge to one of these
minimum points. During initial studies it is necessary to repeat the algorithm with
several starting values and establish that in all cases the algorithm converges to the
same minimum point. When there are more than one local minima all minimum
points should be found and the minimum of these should be taken as the design point.
In some cases the FORM algorithm may fail to converge or values of the basic
variables during iterations may go outside physical limits. In most cases
non-convergence is due to highly non-linear failure surface, in which case, other
algorithms of non-linear constrained minimization should be used.
2-19

2.4.3.2 Second Order Reliability Method

As noted previously, when the failure surface at the design point is highly curved a
tangent hyperplane approximation used in FORM is not satisfactory. In such cases
approximating the failure surface by a quadratic surface, with the same principal
curvatures as the true failure surface at the design point, should give a better estimate
of the probability of failure. This implies a Taylor series expansion of the limit-state
function at the design point in which terms up to second order are retained. Such
methods are called 'second order reliability methods, (SORM)'.

Unfortunately, results for the calculation of the probability content of the failure region
bounded by a general quadratic surface are not available. Good results are available
for hyperparabolic failure boundaries, Madsen et al (1986). For a general failure
surface Breitung (1984) gives the following asymptotic result for the probability of
failure, (detailed derivation can be seen in the reference)

n-I
pp cI(_13) fl(1-3KY'2 -'°° (2.23)
13
i-I

where J3 is the first order reliability index and iq are the principal curvatures of the
failure surface at the design point. The result is asymptotic in the sense J3 - ee with

13 iç fixed. This second order correction to the first order result gives a very good
estimate of the probability of failure for large values of 13 . Tvedt (1988) has derived
a three-term approximation to pf which gives better results even for small to medium
values of 13 . A formal 'second order reliability index' can be defmed as

0 . (2.24)
PSORM W .Pf

FORM and SORM methods discussed so far give reasonably good estimates of the
probability of failure which is adequate in most practical cases. The computational
effort required is very small compared to numerical integration or simulation methods.
2-20
2.5 USEFUL PROBABILITY TRANSFORMATIONS

From the previous discussions it is noted that for reasons of simplicity and the
property of rotational symmetry reliability index and failure probability calculations
are usually carried out in the space of independent, standardized normal variables. If
the basic random variables are initially not in this form, as is nearly always the case,
a transformation to the standard space is necessary. In using the FORM algorithm
described in Section 2.3.3.1, for each iteration, the variables from the standard space
have to be transformed to the original space of basic variables to evaluate the safety
margin and derivatives of the failure surface. In this section some useful probability
transformations, for different initial probability space of basic variables, are presented.

Let, in the following, X = (X i , X2,..., X) be a set of n basic random variables with


an arbitrary joint distribution function F() with the vector of expected values p =
E[X] and the covariance matrix C> = C[X ,XT] or the correlation matrix p, = R[X,XT].
This original space of the basic variables is referred to as x-space. Let U = (U1,
U2,...,U) be a vector of independent, standard normal variables with the expectation
vector E[U] = 0 and covariance matrix C[U, UTI = I, where 0 represents the null
vector and I is the Unit matrix. This space is called u-space. Let T denote the
transformation from x -4 u-space and T' the inverse transformation from u - x-space.

2.5.1 Normal Correlated Variables

In this case the basic variables are already normal. They need to be standardized and
made uncorrelated to obtain U-variables. The transformation T and its inverse P 1 are

T: U - L'D'(X-j1)
(2.25)
T 1 : X-LD(U)+Jt
2-21
where D is a diagonal matrix of standard deviations a of the X-variables and L is a
lower triangular matrix obtained from the correlation matrix such that R = LLT A
lower triangular matrix can be obtained from Choleski decomposition for which
standard algorithms are available, see for example Press et al (1986). Note that for
uncorrelated basic variables R, and hence L, is a unit matrix. The above
transformation can also be used when only second moment information is available
about the basic variables.

2.5.2 Non-normal Uncorrelated Variables

In this case, since the variables are independent, each variable can be transformed to
the u-space separately using the identities

4'(u) - Fx(x1) I- 1,2,...,n (2.26)

The required transformations are


T: U1 -

I- l,2,...,n
(2.27)

T: F(cZ(u)) I - 1,2,...,n

The inverse probability distribution function Fx1() can be determined analytically for
most of the commonly used distribution types, such as lognormal, Gumbel, Weibull,
Rayleigh etc. However, for some distributions such as gamma, beta and student-t
distributions the inverse is to be determined numerically and this can increase the
computational effort of FORMJSORM methods considerably. Rackwitz and Fiessler
(1978) suggest an approximate method in which determination of Fx'() as in Eq.2.27
is not required. At a given point Xj of a variable, the method involves replacing the
distribution FL(xJ) of a variable X1 by a substitute standard normal distribution
function c1-p?)/a']. The parameters ii', a' of the substitute standard normal
distribution are determined such that the values of the distribution functions and the
probability density functions of the actual and approximating functions are identical
at ;. This method is called in Ditlevsen (1979b) as the 'principle of normal-tail
2-22

approximation' and is shown to be applicable to non-normal correlated variables as


well. The use of this approximating method requires some modification to the FORM
iteration algorithm described in Section 2.4.3.1.

2.53 Non-normal Correlated Variables


When the full joint distribution function F() is known, Hohenbichler and Rackwitz
(1981) suggest the use of the Rosenblatt transformation

T: u1 — '[F1(x1)]

(2.28)
U2 — b'[F2(x21x1)]

u,, —

where F1 ( •) is a conditional distribution function given by

ffx1.x2... . x11,x1(i,,,xi_i,t)dt (2.29)


- -
fx1,;... . x,1 (x1,x2,...,x_1)

The inverse transformation can be similarly obtained in a step-wise manner

T: — F[cb(u1)]
(2.30)
x2 — F;1[cb(u2)1x1]

— F;' [cb(u)Ix,,x2,...,x1]

The use of this method, therefore, requires partial conditional distributions instead of
the n-dimensional joint distribution function. Quite frequently complex multivariate
models are constructed in this way. It is observed that the ordering of variables has
some influence on the computed failure probability. Furthermore, the computation of
T ' suffers from the same numerical problems as mentioned before.
2-23

2.5.4 Non-normal Marginals with Known Correlation Matrix

In the case of non-normal correlated variables, sufficient statistical data is not usually
available to construct the joint distribution function or the partial conditional
distributions to use the Rosenblatt transformation described above. In practical
applications, sufficient data may be available to develop (or prescribe) appropriate
marginal distributions for each of the variables and the covariance matrix. A method
of transformation has been suggested by Der Kiureghian and Liu (1986) to deal with
such cases.

A joint probability density function, which is consistent with the known marginal
distributions and correlation structure, is assumed for the basic variables in the form

fx1I)fx2(x2)fxPn) (2.31)
fx(X) - $(y,R')

in which (y,R') is the n-dimensional normal probability density function of zero


means, unit standard deviations and correlation matrix R'. The elements p' of R' are
related to the known marginal densities f,(;) and fx(x) and the correlation
coefficient p of basic variables X 1 and X3 through

=x (2.32)
P12 - j i-i1x2-i2)
a2 f2lY2P 12)t!Y1!Y2
a1
A
The correlation coefficient P12' has to be usually solved iteratively using the above
equation. Simplified empirical formulae, developed from the above relation, are given
in Der Kiureghian and Liu (1986) for different marginal distributions of X1 and X.
For most distributions the difference between p12 and p12' is small.
2-24
Once the correlation matrix R' is determined, the basic variables X = (X 1 , X2,..., X)
with the known marginals F(;) and correlation matrix R, can be first transformed
to a set Y = (Y 1 , Y2 ,..., Y) of variables with n-dimensional normal distribution of
zero means, unit standard deviations and correlation matrix R'. This set can then be
transformed to the required u-space of uncorrelated standard normal variables. The
required transformations are


T1 : Y1 - c [Fx(x)] i- (2.33)


U - L'(Y) (2.34)

The inverse transformation is obtained in a similar way as



T: Y—L(U) (2.35)

(2.36)
7', X - F' [4(Y)] I - 1,2,...,n

It is to be noted that the computation of P12' in Eq. 2.32 is independent of the


linearization point u or x, the correlation matrix R' and hence the lower triangular
matrix L need to be calculated only once before iterations are commenced and there
is no need to determine the joint density function f() explicitly from Eq.2.31.
2-25
2.6 SYSTEM RELIABILITY ANALYSIS

In this section reliability evaluation of Iwo fundamental systems, namely, series


systems and parallel systems is presented. A series system is defined to be the one
in which the system fails when one of its elements fails. On the other hand, in a
parallel system, all elements of the system must fail for failure of the system to occur.
These two systems are fundamental in the sense that the failure of any complex
physical system can be represented in terms of a series system or a parallel system or
a combination of the two. For example, failure of a statically determinate truss
structure can be represented by a series system since the structure fails if any one of
its members fails.

Discussion in this section is limited to the computation of failure probability of series


and parallel systems from the known probabilities of elemental failures. Idealization
of a real complex structure, such as an offshore jacket structure, into a combination
of elementary parallel and series systems is not discussed here. This would require
an understanding of failure behaviour of individual structural elements (such as braces,
joints etc.), load effect re-distribution after elemental failures, time interval between
failure of elements and finally on how the failure event of the structure is defined.
These issues are discussed under relevant chapters in the remainder of this thesis.

2.6.1 Reliability Analysis of Parallel Systems


Failure of a parallel system occurs when all elements in the system fail. The failure
of a parallel system can, therefore, be represented as an intersection of failure events
of individual components. Let there be k elements in a parallel system and let X=
(X1 , X2,..., X) be the set of basic random variables for the system. Let F1 represent
failure event of the tb element and F the system failure event. Then, the probability
of failure of a parallel system is given by

k k k

P1 p - P[F] - P[flF,] - P[fl(M^O)] - P[fl(g(X)^O)]


(2.37)
2-26
th
where M1 = 81(X) is the safety margin and g 1() is the limit-state function for the
element. Thus g1(i)=O defines the failure surface for the ith failure element. As in the
case of an element reliability analysis, exact evaluation of the probability content of
the failure set common to all the componential failure regions with non-linear failure
boundaries is very difficult. First-order and second-order reliability methods as
discussed below can be used to evaluate the probability of failure efficiently with
reasonable accuracy.

2.6.1.1 First-Order Reliability Method for Parallel Systems


In a first-order reliability analysis of parallel systems, the basic variable vector is first
transformed to the space of independent, standard normal variables using one of the
appropriate probability transformations given in section 2.5. For each failure element,
search is then made for the point on the failure surface which is closest from the
origin for which an algorithm of the type given in section 2.4.3.1 can be used. These
points are called 'elemental 3-points'. The failure surfaces of individual elements are
approximated by tangent hyperplanes at their individual n-points. The probability of
failure for the system is then given as, Hohenbichier and Rackwitz (1983)

p1 ,, - P[flg,(X)^O]

- P[flg1'(U)^O]
i-I
k-

_ P[flh(U)^O]
i-i (2.38)
k-
- P[fl(a1U+3)^O]
i-i
k
- P[fl(z1^-f31)]
i-1
-

In the above, 13 are the elemental first-order reliability indices with a1 defining the
vector of direction cosines for the ith failure surface and R is the matrix of correlation
coefficients between the safety margins of pairs of elements, R: = p[M, M] =
for all i^j. Z = aJU is a standard normal variable and c1 k( .) is the k-dimensional
standard normal integral. The only approximation involved in the above is due to the
2-27

linearization of elemental failure surfaces. Calculation of probability of failure of a


parallel system is thus reduced to the evaluation of the standard multi-normal integral.
Unfortunately this is not easy for problems of larger dimensions. A number of
available methods are reviewed by Johnson and Kotz (1972). An approximate but
efficient method has been proposed by Hohenbichler and Rackwitz (1983).

Results of the FORM method will be reasonably good when the componential failure
surfaces are close to linear. However, with increasing non-linearity of the failure
surfaces, FORM results begin to deviate significantly from exact results. Referring
to Fig. 2.2, for example, and focusing attention for the moment on the intersection
failure probability bounded by the two failure surfaces 8i' (ii)=O and 82' (ii)=O, it can
be seen that the probability content bounded by the approximating hyperplanes at the
respective 3-points i.e. h1 (ii)=O and h2(ii)=O can differ significantly from the exact
result. Although FORM results are reasonably good in the evaluation of individual
element reliabilities they become less satisfactory in the calculation of system
reliability. From the discussion to follow it will be clear that poor results from the
FORM approach as above are due to an inappropriate choice of approximation points.

2.6.1.2 Second-Order ReliabiHty Method for Parallel Systems

In the case of element reliability analysis it was argued that the point on the failure
surface closest from the origin (i.e. the design point or n-point) serves as the best
approximating point, for linear or quadratic approximation of the failure surface, as
the probability density of variables decreases rapidly with increasing distance from the
origin and thus only the near vicinity of the design point contributes significantly to
the integral in Eq. 2.4 for p 1. Using results of asymptotic analysis, Hohenbichier et
al (1987) have shown that such an argument also applies to more complex failure
domains formed by the intersection of several componential failure domains. A
distinction is therefore made between 'local p-points' as related to individual
elemental failure domains and the 'joint f3-point' Ii,'' for the parallel system given by
the intersection of elemental failure events.
2-28

Li

Fig. 2.2: Probability of failure calculation for a parallel system using SORM.

Thus in Fig.2.2, points i, r, ii and i are local f3-points and i12*, and jj344
represent joint 3-points for the intersection events (F1 11F2), (F2flF3) and (F31tF4) while
= ii is the global joint 3-point for the parallel system (F1 flF2flF3flF4). It can be
visualized that linearization of failure surfaces g()=O and g(ii)=O at their joint
3-point 1? gives a better approximation for the probability of failure P(F1flF2).
Similarly the joint u-point i1 is a better approximating point for the failure domain
of the parallel system (F1 flF2flF3flF4). For a parallel system with k elements, the joint
3-point ii can be determined from a constrained minimization of the form

- minlul
(2.39)
with g'(i) ^ 0 for i - l,2,...,k

A number of algorithms for the non-linear multiple-constrained minimization suitable


for this task are available, see forexample Gil1, Murray and Wright (1981). An
algorithm given by Schittkowski (1986) is used in the program RASOS (Reliability
Analysis Software for Offshore Structures) developed under the BR1TE Project, see
RASOS-R theoretical manual (1991). A detailed presentation of this algorithm is
2-29
considered beyond the scope of this thesis. It is to be noted that in Eq.2.39 all the
constraints are not necessarily active i.e. g(ii)=O is not necessarily true for all k
elements at the joint j3-point. Let the constraints be numbered such that the
constraints 1 ^ i ^I are 'active' and +1 ^ i ^ k are 'inactive' constraints.

Considering joint 13-point as the approximating point, Hohenbichler and Rackwitz


(1987) have derived the following second-order asymptotic result for the probability
of failure of a parallel system. At the joint 13-point ir, unit normal vectors with
respect to each active elemental failure surfaces are calculated as


a1 - (2.40)
IVg1'(i*)I

It is assumed that the vectors a are all linearly independent, implying that I ^ n, the
dimension of random variable vector. Since u is given as a minimum point from
Eq.2.39, there exist constants y1 : ' ^ 0 for 1 ^ I I which can be uniquely
determined by solving the equations

1 1

- E
i-I
yVg1'(*) - E
i-i
y1a1IVg1*)I (2.41)

Assuming further that the failure functions g( .) are twice differentiable at 11, a matrix
D containing the second order derivatives of the active constraints is defined as

- I a2g(u*)
D: d 11 - for l+1^i ; j^n (2.42)
Ey

The probability of failure for the parallel system is then given by

P1,p - .b, (-6 ; R) [1-D ) j-.l /2 •_ (2.43)


2-30

- (2.44)
8: 8 1 -a1 u* for 1-1,2,...,!

- a1Ta for i,j - 1,2,...J (2.45)

The -dimensiona1 standard normal integral can be evaluated using the method given
by Hohenbichler and Rackwitz (1983).

Comparing Eq.2.43 with Eq.2.38 it can be seen that the first term in Eq.2.43 amounts
to linearization of failure surfaces of active elements at the joint f3-point. This
procedure can be called 'improved FORM' method. Often this itself gives
considerable improvement over the 'crude FORM' method presented in section 2.6.1.1.
The second term in Eq.2.43 represents the second order correction to account for the
curvatures of the failure surfaces at ii'. In the special case of a system with single
element Eq.2.43 reduces to the SORM result for element reliability given by Eq.2.23.
Furthermore, it is to be noted that in Eq.2.43 contribution of only the active failure
elements is considered. The contribution of inactive elements can be neglected
asymptotically for large I . Alternatively, they can be included in Eq.2.43 by
enlarging the dimension from Q - k. The required gradients for the inactive elements
can be approximately calculated at the joint j3-point 1:?, or if necessary, 'higher-order
joint p-points' (e.g. points 1:i*, I1 3. in Fig.2.2) can be determined as in Eq.2.39.

From the foregoing it appears that an estimate of the probability of failure for a
parallel system can be improved in stages as given below:


I Individual linearization, use elemental FORM in Eq.2.38.

II Individual linearization, second order corrections computed at individual
[3-points, use elemental I3SORM in Eq.2.38.

ifi linearization at the joint [3-point, no second order corrections, only active
elements included, p from the first part of Eq.2.43.

Iv Same as iii but include second order corrections in Eq 2.43.

V Same as IV but also include contribution of inactive elements with gradients
evaluated at the joint [3-point.
2-31

VI Same as V but evaluate gradients for inactive elements at their higher-order


joint 3-points.

It is not clear what terminology should be used to refer to the above procedures.

2.6.1.3 Probability Bounds for Parallel Systems

It is obvious that reliability of a parallel system increases with increasing number of


its components. For a given system, with fixed number of elements, the system
reliability decreases with increasing correlation between pairs of elements. Thus by
considering that component failure events are fully correlated an upper bound to the
system probability of failure can be obtained. Analogously zero correlation between
pairs of events corresponds to a lower bound on p (i.e. upper bound on J3). These
bounds can be expressed as,Thoft-Christensen and Murotsu (1986)

k k k
HP(F1 ) <P( flF a ) ^ minP(F1) (2.46)
i-I i-I i-I

These bounds, however, become too wide to be of any value for moderate correlation
between failure events. Unfortunately, efficient methods are not available to decrease
the gap between these bounds. In Section 2.6.2.3 narrow bounds are given for the
probability of 'union of failure events' (series system). In principle, since
P(fl M^O)=1-P(U M>O), bounds for the union of events can be used to determine
bounds for a parallel system. However, the quality of these bounds are seen to be
rather poor for high probability events and therefore this approach cannot be used
successfully for high reliability structural components.
2-32
2.6.2 Reliability Analysis of Series Systems

A series system fails if any one of its elements fails. Probabilisticaily, therefore,
failure event of a series system with m failure elements can be represented as the
union of failure events of individual components. The probability of failure of a series
system is given by

P[U(g'(X)^O)] (2.47)
Pj, 5 - P[F5 ] - P[UF1 ] -

The corresponding reliability index for a series system is

(2.48)
I3 s -

The reliability of a series system can be computed using first order or second order
reliability methods as discussed below.

2.6.2.1 First Order Reliability Method for Series Systems


A first order method for series systems involves linearization of elemental failure
surfaces at their respective [3-points. The probability of failure of a series system is
given by, see Hohenbichier and Rackwitz (1983)

P1,s - P[Ug(X)^O]

- P[Ug1(U)^O]

P[Uh (U) ^O]

- P[U(a1U+[31)^O] (2.49)

- P[U(z^-[31)I

- 1-P[U(z1>-[3)]

- 1-m(3;R)
2-33
where ; = U is a standard normal variable, f3 is the vector of elemental reliability
indices and R is the correlation matrix of componential safety margins with elements

2.6.2.2 Second Order Reliability Method for Series Systems

In the case of series systems, the failure set for the system is made up of the union
of all componential failure sets. Thus eveiy point in the system failure set is
contained in at least one of the componential failure domains. The local n-points of
all componential failure surfaces are also the local f3-points of the system failure
domain. The 'global 3-point', or the point of highest probability density on the
system failure boundary can be determined simply by (3G10 = mm , say J3,
corresponding to the (" component. Using global n-point as the approximation point,
the system failure probability will be close to the probability of failure of the weakest
element (with lowest t 3 or highest p) of the system. However, since the contribution
of the remaining elements is neglected this will underestimate the true system failure
probability. The contribution of the remaining components can be included by simply
approximating their individual failure surfaces at their individual 3-points (since these
axe also the local [3-points of the system failure boundary). A linear approximation
will lead to Eq.2.49 by the FORM approach described above. In a second order
reliability method, approximation of individual failure surfaces by quadratic surfaces
at the local [3-points amounts to using [3i,SORM in the last step of Eq.2.49.

2.6.2.3 Probability Bounds for Series Systems

Reliability of a series system decreases with increasing number of its elements. For
a given system, with fixed number of components, system reliability decreases with
decreasing correlation between pairs of elements. Simple bounds on the probability
of failure of series systems can be constructed by arguing that zero correlation
between any pair of elements gives an upper bound on the probability of failure (i.e.
lower bound on [3) while full correlation corresponds to lower bound on p (i.e. upper
bound on I3 ). The simple bounds are, Thoft-Christensen and Murotsu (1986)
2-34


PFI^PI,S^l -rI[1-P(F1)] (230)

However, these bounds become uninterestingly wide in the practical cases of


intermediate dependencies. Narrow bounds on p, can be obtained by using the
expansion theorem for the probability of the union of events, i.e.

P[UF1] - i: (-1)Sk (2.51)

Sk E P(F1flFfl...flF)

The partial sums in the above equation alternate around the true value, S 1 ^ p,
S2 ^ p, S3 ^ p f ••• . Unfortunately it is not true that the upper bound decreases and
the lower bound increases with k. Moreover, it involves the evaluation of an ever
increasing number of intersections. Ditlevsen (1979a) has derived the following
bounds by considering the intersections of only any two failure events.


^ P(F1 ) ^ max[0 , P(Fi)_EP(FiflF1)]

(2.52)

^ P(F1) (P(F1 ) - max(P(FflFj)))


+

The bounds depend on the ordering of the failure events which can be different for
the upper and lower bound. Algorithm for an optimal ordering has been suggested in
Ditlevsen (1979a). These bounds are very narrow for near full dependencies or near
independence between failure events. Their quality is the worst for medium
correlations between failure events and if the correlations are the same between any
of the failure events and the events have the same probability. The quality increases
with decreasing componential failure probabilities. The difference between the bounds
increases quadratically with the number of components, Hohenbichier et a! (1983).
2-35

2.6.3 Reliability Analysis of General Systems

In a general redundant system with a large number of components failure of the


system occurs after a certain number (more than one) of elements have failed. The
failure event of such a system can be represented in terms of 'minimal cut sets' or
'minimal tie sets' of the system and the system reliability calculated as below.

A 'cut set' is defined as a set of elements which, if they all fail, results in system
failure. A cut set is said to be 'minimal' if no subset of elements from the cut set can
be found which would result in the failure of the system. Clearly for a system with
large number of components several minimal cut sets can be found and system failure
occurs due to failure of any of these minimal cut sets. The probability of failure of
the system is then given by

m k.
(2.53)
Pj, - P[U(fl(Mfr(X)^0))]
j.. r-1

where X is the basic random variable vector for the system and M, is the safety
margin for the rth element in the 1th minimal cut set which contains k1 number of
elements and m is the total number of possible minimal cut sets for the system.

A tie set (path set or link set), on the other hand, is defined as a set of elements
which, if they all survive, assures the survival of the system. A minimal tie set is
defined as a set of minimum number of elements which must survive for the survival
of the system. Failure of any element in a minimal tie set results in system failure
and thus it represents a series arrangement of elements. In a system with large
number of elements the whole system survives if at least one of the minimal series tie
set survives. The system can therefore be seen as a parallel arrangement of the
minimal series tie sets the probability of failure of which is given by

: I, -

Pf.jys - P[fl(U(M(X)^O)] (2.54)


J r-I
2-3O
where M. is the safety margin for the n1h element in the th minimal tie set which
contains ç number of elements and s is the total number of possible tie sets for the
system. The system failure events represented by Eq. 2.53 and 2.54 are equivalent.

Reliability of general systems can only be evaluated in terms of upper and lower
bounds on the probability of failure given by Eqs.2.53 or 2.54. For the minimal cut
set representation, close bounds can be obtained by using the Ditlevsen bounds for
union of events given in Section 2.6.2.3 for pure series systems, in which each event
F1 is now thought of as an intersection of further (componential) events Ffr, i.e.
Fj=flF.. In Eq.2.52, P(F,) now corresponds to a cut set probability which can be
evaluated by one of the methods discussed for pure parallel systems; while, P(F1I1FJ)
terms can be similarly evaluated by increasing the number of components in the
parallel system to (k1 ^k). In order to use these bounds for the minimal tie set
representation of systems one has to use complements of the events given in Eq.2.54
(i.e. survival events) and for high reliability components these become events of high
probability for which Ditlevsen bounds are unsatisfactory. This is one of the reasons
why minimal tie set model is not often used in system reliability analysis.

An alternative approach to the use of bounds is possible if each intersection (Cut set)
in Eq.2.53 is replaced by an equivalent event with the same probability. Gollwitzer
and Rackwitz (1983) have suggested a method for obtaining an 'equivalent linear
safety margin', MC for a parallel system which defines an 'equivalent hyperplane'
which cuts off the same probability content as that of the probability of failure for the
parallel system. In terms of these equivalent linear safety margins for each cut set the
probability of failure for the system in Eq.2.53 can be re-written as

'Pt
(2.55)
pf,3ys - P[U(M1'P(X)^Ofl
i.1
2-37
The probability of this pure series system can now be bounded using Ditlevsen bounds
or evaluated directly as in Eq.2.49 using the rn-dimensional standard normal integral.
This method is computationally more efficient if the total number of basic variables
is less than the number of cut sets (i.e. n^m^30), Gollwitzer and Rackwitz (1988).

2.7 SENSITIVITY AND IMPORTANCE MEASURES

In structural reliability analysis, in addition to the value of the probability of failure


or reliability, it is often of interest to know which of the basic random variables is
stochastically more important in its contribution to the probability of failure.
Similarly, when the input information about the distribution parameters of a variable
are not precise it is useful to know the sensitivity of the probability of failure to
changes in distribution parameters. This will help to decide for which variable more
statistical data is to be collected or rigorous quality control is to be exercised during
construction. In a system reliability analysis it is of interest to know the "criticality"
of each component to system survival or the "importance" of each cut set to system
failure. A number of sensitivity and importance measures have been reported in the
literature which provide valuable additional information to help in decision maldng
under uncertainty and these are discussed in the following.

2.7.1 Direction Cosine Sensitivity Factors

chich
In Section 2.4.3.1 a vector of direction cosines represents a unit vector normal to
Int-odAtce4
the failure surface at the design point u From Eq.2.20, defining 1 3 as the distance
from the origin to u it directly follows that
2-38

a1 (2.56)

The numerical value of cx1 is thus a measure of the sensitivity of the reliability index
to inaccuracies in the value of u j at the design point. Thus a5 have often been referred
to as 'sensitivity factors' in reliability literature, see Madsen, Krenk and Lind (1986).

Corresponding to Eq. 2.19a for the linearized failure surface, an expression for
linearized safety margin in the standard normal space can be written as

(2.57)
M, -

From this the variance of M1 can be obtained as


Var(M,) - Ea - 1 (2.58)

and c4 can also be interpreted as the fraction of the total uncertainty caused by the
uncertainty described through U 1. In the case of independent basic variables the
uncertainty described through U 1 is directly related to the uncertainty in X 1 and hence
in this case axlOO can be interpreted as the percentage importance of the basic
variable X1 in its contribution to the total uncertainty. Reduction of uncertainty (by
more data collection or strict quality control) of a variable with higher a will result
in significant increase in the reliability of a structure or a component. However, when
the basic variables are correlated the transformation Eq.2.27 shows that U depends on
more than one variable and in this case, therefore, a cannot be interpreted as the
contribution of X1 to total uncertainty.

2.7.2 Omission Sensitivity Factors

In order to determine the relative error in reliability index f3 by treating one of the
basic variables as deterministic Madsen (1988) introduces an 'omission sensitivity
factor'. For a basic variable X3 ,it is defined as the inverse ratio of the first-order
2-39
reliability index 13 and the first order reliability index f3' with Xi replaced by a
deterministic value cia. For the general case of non-normal dependent variables
omission sensitivity factors for each variable can be obtained as

1 [(J.I)T] [(1* -Ji*) 1 - (2.59)


[1_TJi(JJT)J(J-I)T]I/2
- 13

For normal correlated variables and the transformation to u-space defined as in


Eq.2.24, this simplifies to


- 1+(LTDc)1 (.L-q1)/j3
(2.60)

which for independent basic variables gives finally

1- cU Uq./ 13 _____
Yj - ________ withu - 0.
(2.61)

In the above J is the Jacobean matrix of derivatives J 1 = evaluated at the


design point, D is a diagonal matrix of standard deviations o, L is the lower
triangular matrix of correlation coefficients such that R = LLr and G is the matrix
with elements G, = pcYcY etc. in the jthl row and th column and zeros elsewhere. Ji
is similarly defined to G with elements J. The results are asymptotically exact for
large 13 and it is often convenient to use the mean value of the variable tj for q.

For independent variables the relative error in the reliability index by replacing a basic
variable by its mean value is found to be less than 1% for a sensitivity factor with
absolute value less than 0.14. For a specified accuracy of the reliability index the
number of basic variables in a probabilistic model can often be reduced based on their
omission sensitivity factors resulting in computational savings. Since an omission
sensitivity factor quantifies the change in reliability index by reducing the variability
in a basic variable to zero it provides a better measure of the stochastic importance
of the variable.
2-40

2.7.3 Parametric Sensitivity Factors

When the statistical information about basic variables is limited and hence the
estimated distribution parameters are imprecise, it is of considerable interest to know
the sensitivity of the failure probability or the reliability index to small variations in
the parameters. It is also of interest to determine the sensitivity of the reliability index
to changes in some deterministic input parameter in the limit state function. Let, in
the following, i denote a vector of parameters which includes both distribution
parameters and deterministic quantities. In the
following asymptotic approximations to &[RThti are given which do not require a
repeated computation of the reliability index.

From Section 2.4.3.2 and using the asymptotic, second-order approximation of


Breitung (1984) it is clear that PR PSORM PFORM = 3 for 3 -, . Therefore,
asymptotically, aPR/a;=/tI. Consider the reliability analysis of a single element
and let ( 0) be the computed first order reliability index evaluated corresponding to
the initial parameter vector . It is clear that in the formulation for j3 in the standard
normal space, a deterministic parameter must enter the formulation through the
limit-state function. Re-working Eq.2. 16 for the failure surface in the u-space

- g(T(;i0 )) - g(; 0 ) (2.62)

The distribution parameters, on the other hand, enter the formulation through the
transformation T in the above which can be obtained as given in section 2.5.
Re-writing Eq.2.20 for the reliability index

(2.63)
- I *I -

The parametric sensitivity factor for a distribution parameter is then given by,
Hohenbichier and Rackwitz (1986) and Madsen et al (1986)
2-41

a R(0) *
T(.; 0)
_______
ati
-
1—u* rau - !T a (2.64)

This requires only a derivative of the transformation T with respect to ç. Similarly


for a deterministic limit-state function parameter, the result is

a- .u* ._u*
.43R(tO)

11 Vg(i*;10)Tl a -
;*)j]14*
lVg(* (2.65)

;•)

- IVg(i*;0)I

For a parallel system based on Hohenbichier et al (1987) it has been shown in Section
2.6.1.2, that the joint p-point is given by Eq. 2.39 and the parametric sensitivity for
a deterministic parameter can be given by

- 1 - a -
at0) -_u* _u*

a -
- us (2.66)
f3j-i

and for a distribution parameter it is


2-42

-
atI3R(t0) 1 - TaT(*;) (2.67)
ati
In the above 13* is the length of u or the distance of the joint 13-point from the origin.

The parametric sensitivity of the probability of failure is obtained from the sensitivity
of the reliability index through

- --(-13R(o))


— (2.68)

a-

Finally for a pure series system, assuming that the failure probability is asymptotically
equal to the sum of the component failure probabilities, the sensitivity factors of the
system probability of failure are

m
() P1(t)
?
(2.69)
in
a -

i-I

These parametric sensitivity factors are of considerable practical importance in the


optimization of structural design and in reliability updating. Similar parametric
sensitivity for direction cosines have been derived by Bjerager and Krenk (1989).
2-43
2.8 CONCLUDING REMARKS

This chapter has presented the basic theory of reliability analysis of structures.
Probabilistic basis of structural safety is discussed and the probability of survival or
reliability is introduced as a rational measure of structural safety. More specifically,
the meaning of the safety measure is clarified as a theoretical value of reliability
computed from a reliability model and its usefulness in rational decision making under
uncertainty is discussed. Various methods for reliability analysis of individual
structural elements are presented and in particular the use of FORM and SORM
methods is advocated in practical applications for their efficiency. First order and
second order methods axe also presented for reliability analysis of parallel systems,
series systems and general systems. In particular it is noted that for the minimal cut
set representation of systems, the probability of individual cut sets can be evaluated
based on the multi-normal integral while the probability of the final union event can
be evaluated using either Ditlevsen bounds or equivalent hyperplane approach. This,
however, requires that all the cut sets for the system are identified in advance.

The idealization of a real complex structure, such as an offshore jacket structure, into
a minimal cut set system requires a careful consideration of failure behaviour of
individual structural elements (such as braces, joints, piles etc.), load effect
re-distribution after elemental failures, time interval between failure of elements and
changes in both the loading and strength of members during this time and finally, and
most importantly, the definition of failure event for the structure. Each minimal cut
set corresponds to a 'failure mode' of the structure and for a structure with large
number of elements and high degree of redundancy there are usually a large number
of possible failure modes and it is generally not possible to identify all of them. Even
if all failure modes are identified it will be computationally very expensive to include
all these cut sets in system reliability evaluation. Therefore a suitable strategy is
required to identify only "stochastically dominant" fai'uire modes which contribute
most to the system probability of failure.
2-44
System reliability analysis of offshore jacket structures under extreme load and fatigue
conditions, therefore, involves the following broad steps:

(1) Develop appropriate mechanical models (deterministic or probabilistic) to


determine response of jacket structures to environmental loading in terms of
member forces or stresses.

(2) Develop suitable models to determine the capacity of individual members and
joints under different failure modes and also to determine the ultimate strength
of the structure.

(3) Identify major sources of uncertainty, due to random loads and material
resistance properties and also in the mechanical models used for response
prediction and strength estimation. Quantify these uncertainties through a set
of basic variables with appropriate probability distributions. Carry out
reliability analysis for each failure mode of individual members and joints.

(4) With an appropriate definition of structural failure identify all stochastically


dominant failure modes (minimal Cut sets) using a suitable search strategy.
Evaluate the probability of failure of the minimal cut set idealization of the
structure.

Step 1 is the subject of discussion in Chapter 3 where response modelling under


extreme wave load and long-term fatigue loading is considered. Deterministic models
for the capacity of tubular joints failing by fatigue, fracture or plastic collapse are
presented in Chapter 4. The behaviour of bracing members is not considered in detail
as the emphasis of this thesis is on fatigue deterioration at tubular joints and its
influence on system reliability.
2-45
Chapter 5 presents a detailed discussion on the sources of uncertainty involved in the
failure prediction of tubular joints and appropriate probabilistic models are suggested
for quantification of these uncertainties. Reliability analysis is carried out for the
limit-states of fatigue failure, fatigue-fracture and fatigue-plastic collapse.

System reliability evaluation of a jacket structure is presented in Chapter 6. Here,


system failure is defmed by the collapse of a structure under a single (life-time
extreme) wave. A modified algorithm for non-linear structural analysis for a
prescribed failure sequence is presented and a modification of the "branch-and-bound"
method is given for the search of stochastically dominant failure modes. A simplified
approach is also suggested for including fatigue deterioration in the above analysis.
A time-dependent system reliability method is developed in Chapter 7 for analysis of
pure fatigue/fracture failure sequences and a combination of extreme load and fatigue
failures within a single failure-tree.
Chapter 3
MODELS FOR LOADING AND RESPONSE OF OFFSHORE STRUCTURES

3.1 GENERAL

As mentioned in Chapter 1 the loading on offshore jacket structures consists of


man-made loads, arising from the self-weight of the jacket and weight of the deck,
and those imposed by the ocean environment. However, the biggest source of
uncertainty in the design or assessment of an offshore structure occurs in the
esthnation of extreme environmental conditions and the computation of the response
of the structure to environmental loading. in a conventional deterministic approach
the difficulty has always been in the choice of an appropriate set of environmental
conditions to be used for the design of these structures. In a reliability analysis the
environmental conditions which may occur within the service life of a structure need
to be represented in terms of appropriate stochastic models.

In order to determine the environmental loading and response of jacket structures, two
methods are, at present, in common use. The first is a static, discrete wave
deterministic analysis which considers the action of a single wave, described by the
wave height and wave period, and associated uniform wind and current conditions,
described by their velocities. The second is a stochastic approach which considers the
action of a number of waves by a suitable representation of the temporal variation of
the sea surface over a short period. The analysis can be carried out either in a
time-domain or a spectral (frequency) domain and the approach is usually
semi-probabilistic. In principle, both these response analysis methods can be used in
a reliability assessment of offshore structures, which thereby need different forms of
probabilistic representation of the environmental conditions. The response of the
structure obtained from these two methods will also be in different forms and suitable
probabilistic models are required to characterize the randomness in response in the two
cases. For the ultimate strength design of a structure the interest is usually in the
"extreme" response expected during the life-time of the structure whereas for fatigue
assessment a long-term response description is required.
3-2

In the present chapter, probabilistic models required in a reliability analysis to


characterize the inherent randomness in environmental action are discussed while the
modelling of other sources of uncertainties in loading, such as uncertain extent of
marine growth or the uncertainties in the methods used for response estimation, is
presented partly in chapters 5 and 6. The emphasis of discussion in the present
chapter is on the structure's response to wave action which forms the predominant part
of environmental loading in extreme response estimation and is the only form of
cyclic loading which causes significant fatigue damage. Probabilistic modelling of the
environment is presented under Section 3.2 for both short-term and long-term
conditions. The two methods of response analysis are summarized in Section 3.3
while probabilistic description of the response obtained from these two methods is
discussed in Section 3.4 with regard to extreme response modelling. Section 3.5
presents, in detail, the modelling of long-term cyclic stress range distributions required
for fatigue assessments. Presently available methods for determination of cumulative
fatigue loading are reviewed and two new probability density functions are developed
to model variation of stress ranges in a short-term and in the long-term. Methods to
account for stress endurance or threshold effects in the evaluation of cumulative
fatigue loading are given.

3.2 PROBABILISTIC DESCRIPTION OF OCEAN ENVIRONMENT

Environmental loading on offshore structures comprises the action of wind, waves and
currents all of which vary appreciably with time. For most shallow-water jacket
structures wave loading is dominant and this section is presented with special
emphasis on waves. A joint probabilistic description of wind, wave and current is
given in Section 3.2.4.

Waves are caused by the action of wind over the sea surface and can be described in
terms of the variations in the instantaneous water surface elevation above mean still
water level, denoted tl(t). Alternatively, they can be represented in terms of a 'wave
height', H, defined as the total range of r(t), in the time interval T between two
consecutive zero up-crossings by 11(t), called 'wave period', for each individual wave.
3-3
The sea surface elevation changes continuously in time and hence q(t) can be
considered as a stochastic process. Since mean still water level is taken as the origin,
11(t) is described as a zero-mean, continuous space, continuous-time stochastic process,
the statistical properties of which can be represented in terms of its 'auto-correlation
function', R,m(t). Observations show that statistical properties of q(t) do not change
appreciably over short durations, typically 3-6 hours, such a sea condition being called
a 'sea-state', over which 11(t) can be considered as a 'stationary process'. Therefore,
for statistical and mathematical convenience, the wave environment at an offshore
location is represented in terms of a probabilistic model of tl(t) for a short-term sea
state and in terms of the variation of the parameters of this model to represent the
long-term variation of 11(t). Analogous to 11(t), waves can also be represented in terms
of probabilistic models for the short-term and long-term joint variation of wave height
and wave period of individual waves.

A long-term description of wave climate is generally of interest in response analysis


required for cumulative phenomenon such as fatigue. However, for ultimate strength
design or assessment of offshore structures, what is normally required is a probabilistic
description of the "extreme" wave condition that can occur once in the life-time of a
structure. In the following probabilistic models for short-term, long-term and extreme
wave conditions are presented.

3.2.1 Short-term Wave Climate Modelling:

As mentioned previously, over a short duration, the sea surface elevation at a specific
position in the sea can be modelled as a zero-mean, stationary stochastic process. For
mathematical simplification this process is often assumed to be Gaussian and
narrow-banded, see for example Sarpakaya and Issacson (1981). The Gaussian
assumption allows the sea surface to be expressed as an infmite sum of independent
sinusoidal wave components. The narrow-band assumption implies a symmetric
process with each crest followed by a trough of the same amplitude. Such an
assumption is reasonable for small amplitude waves. For steep waves, however, crest
3-4
heights are observed to be larger than troughs, Sarpakaya and Issacson (1981). In this
case the distribution of 11(t) is skew and the Gaussian assumption is unrealistic.

Alternative to the above time-domain representation of Tl(t), the sea surface elevation
can also be represented in terms of power-spectrum in the frequency domain. In
practical applications it is most commonly represented by a one-dimensional 'spectral
density function', G,(o), which is a complex Fourier transform of the
auto-correlation function, R(t) of 1(t) in the time-domain. A number of practical
forms of spectra have been suggested for ocean waves of which the
Pierson-Moskowitz spectrum (P-M) is most widely used. This and most other
one-dimensional spectra are characterized by the two statistical parameters namely:
the 'significant wave height', H1, defined as the mean height of the highest one-third
of all the waves recorded during a small duration of observation, and the 'mean
zero-crossing period', T, defined as the mean of sequence of times T between
successive up-crossings of the still water level during the same period of observation.
In terms of these parameters, the Pierson-Moskowitz spectrum can be given as

G(o) - Aoexp[Bo]
2(
A LI
4i (3.1)
"4
B-!.?.
itT2

where the exponents and are normally assumed to be 5 and 4 respectively. The
JONSWAP spectrum used for the North Sea environment can be obtained from the
P-M spectrum by using a correction factor for limited fetch conditions.

The P-M spectrum as given above describes the distribution of wave energy about a
mean direction, 0 0 . In order to describe the spread of wave energy about the mean
direction, as in real seas, a directional spectrum G(o,0) is required. A directional
spectrum is often represented in terms of an uni-directional spectrum as
3-5

G(w,O) - G(w)d(w,8) (3.2)

where d(w,9) is a 'directional spreading function' which depends both on frequency


and direction. However, in practical applications enough data are usually not available
to determine this function accurately and a spreading function, dependent only on
direction, is assumed arbitrarily, Sarpakaya & Issacson (1981).

For a particular sea-state, waves can also be described in terms of probability


distributions of wave height and wave period. It can be easily shown that for a
narrow-banded, Gaussian process q(t), wave heights follow a Rayleigh distribution.
Since wave height and wave period are correlated what is required is a joint
distribution of the two. For a relatively narrow-banded process and assuming
stationarity and Gaussian conditions Longuet-Higgins (1983) gives the following result

f(p,#r) - 2L(e) -_
p2 ex[_2(1 +(1_!)2c2J] (3.3)

where p and t are dimensionless parameters of wave amplitude a and period T

p - a/,f2m0 H12,/2m0

2E m0

m - £ e " G,( o ) do n - 0,1,2 etc.

mç/n2 1/2
C -r 2
[m

1 - 1
L(c) ••

where C is a 'spectral band-width parameter' and L(e) is a normalization factor.

As shown by Turner and Baker (1988) this joint-distribution can be represented in


terms of a marginal density function of wave amplitude f(p) and a conditional
distribution of wave period f(r p). Integrating Eq.3.3 over all possible values of t
3-6

f(p,'r)dt - 2L(c)p(p/c)exp(-p2) (3.4)


1(r)

where (1(') is the standard normal distribution function. The conditional density and
distribution functions of 'r are given by
(3.5)
f(tlp) - f(p,t)If(p)

__ - 1 4P(1!)] (3.6)
F(tlp) - f(P)s;f(P)dt
4(p'/e)

This representation is more convenient in a reliability analysis using the Rosenblatt


transformation, see Section 2.5.

3.2.2 Long-term Wave Climate Modelling

Over a much longer time-scale the sea-surface elevation process r(t) can no longer
be considered as a stationary process. Since the theory of non-stationary stochastic
processes is not well developed analytical models for the long-term variation of the
sea-surface elevation process are not available. However, any longer time interval can
be conveniently considered as a sequence of short-term intervals over which r)(t) can
be described as a stationary process characterized by the short-term statistical
parameters, significant wave height (I-I s) and mean-zero crossing period (Ti). Thus
long-term wave climate can be modelled in terms of long-term variations of H and
T. For response analysis, ideally, a long-term joint-probability density function of
significant wave height, mean zero-crossing period and mean wave approach direction
f(H,T,O), is required. Again because of non-stationarity this long-term distribution
cannot be derived analytically and has to be developed from actual offshore
measurements.

A variety of devices are used for recording waves at an offshore location. Waves are
recorded intermittently, with a recording period of 10-20 minutes and at recording
3-7

intervals of 3-6 hours. Most wave data gathering programmes record only wave
height and period and wave direction is not measured simultaneously. Because of this
a joint distribution of all the three variables is not available. For each wave record
the significant wave height and zero-crossing period are determined. Wave data can
be presented in different formats, see Sarpakaya and Issacson (1981), but the most
common representation is the so-called 'wave scatter diagram' which gives the
proportion of time for which each pair of H 1 and T were observed. This
representation, unfortunately does not indicate any seasonal fluctuations and all
sequential information is lost. A scatter diagram is usually developed from data
extending only over a few years and as a composite of all wave directions and is
supposed to represent wave conditions in any typical year, sometimes referred to as
'annual scatter diagram'. Thus a scatter diagram gives a joint-probability density
function of only H 1 and T2 in a discrete form. A probability distribution for wave
direction (again in a discrete form) is usually obtained from wind measurements
assuming it to be independent of H 1 and Tr

For fatigue damage assessments an annual scatter diagram is usually adequate as most
damage is caused by moderate waves in most frequently occurring sea-states.
However, for extreme response predictions a scatter diagram representative of wave
conditions over the proposed life-span of a structure is required. Inglis et al (1985)
suggest that a notional 100-year scatter diagram' or preferably '1000-year scatter
diagram' should be used. In order to extrapolate wave data obtained over only a few
years to a much longer time-span, mathematical models for H 1 and T are required.
A number of models such as lognormal, Gumbel or Weibull have been used to fit data
on significant wave heights, see for example Sarpakaya and Issacson (1981), which
gives a marginal distribution of H1 . The conditional distribution ofT 2 given H1 is well
modelled as lognormal, see Haver and Winterstein (1990). Inglis et al (1985)
describes the procedure for extrapolation of H 1 and T to develop scatter diagrams for
longer reference periods which also gives a '1000-year scatter diagram' for Northern
North sea conditions.
3-8
Similar to H and T long-term distributions can be developed for individual wave
heights H or maximum wave heights, Hm, in each sea-state and the associated wave
period T or Tm. When available wave data is inadequate, specially for the prediction
of extremes, wave conditions are sometimes inferred from wind data, which are
generally more extensive, by a procedure called 'wave hindcasting'. This procedure
is especially useful for hurricane generated waves such as in the Gulf of Mexico.

3.2.3 Extreme Wave Condition Modelling

Extreme wave predictions have been of considerable interest in deterministic design


practice of offshore structures based on a 'design storm' or 'design wave' approach.
Design wave conditions are chosen from considerations of 'return period', 'wave
exceedance probability' or 'wave encounter probability', see Sarpakaya and Issacson
(1981). In the design storm approach, the data on significant wave heights is
extrapolated using a suitable probabilistic model for H (for example, Gumbel or
Weibull) to determine a design significant wave height, HS,d with a specified return
period or encounter probability. In the design wave approach, the collected data on
maximum wave height in each sea-state is extrapolated to obtain the 'design wave
height'. The associated periods (TZ,d or Td) axe obtained, often using some arbitrary
wave steepness with limitations of wave breaking. However, the extrapolation as
above leads to a point estimate of the mean value of the relevant wave height. In a
reliability analysis the interest is on the joint probability distribution of the most
extreme (maximum) sea-state, f(H, TZ)E, that can occur in the proposed life span of
the structure and this can be obtained as follows.

In a storm-based approach for response analysis, the probability distribution of the


life-time extreme significant wave height, (H)E, may be derived from the data on
annual extremes using the asymptotic theory of extreme-value distributions. Turner
and Baker (1988) have modelled the life-time extreme significant wave height as
Gumbel distributed the parameters of which are obtained analytically from a Gumbel
model of the annual maximum significant wave height. In order to estimate these
parameters with confidence wave data extending over several years is required. In the
3-9
absence of such data wave hindcast methods from wind data may be used. Significant
statistical uncertainties exist in the prediction of extremes and must be considered in
a reliability analysis. The distribution of the associated zero-crossing period for the
extreme sea-state, r conditional on (HJ)E is often modelled as lognormal, Haver
and Winterstein (1990). Alternatively, Turner and Baker (1988) suggest modelling
'characteristic wave steepness', defined by Q = (2itH)/(gT 2) as a lognormal variable
independent of H which seems to be a satisfactory assumption for severe sea-states.

In a discrete wave approach for response analysis, the probability distribution of wave
height, (H)E of the most-extreme wave can be derived conditional on the parameters
(11, TZ)E of the most extreme sea-state. Using order statistics the distribution FXm(X.)
of the most extreme value Xm out of N trials of a random variable X with a
distribution function Fx(x) is given by

F (Xm) - [Fx (x)} N - (1 - [1 - F1 (x) ] ) N (37)

f(Xm) - i [F(X)]N -
N [Fx(x)]'fx(x) (3.8)

Considering the sea-surface elevation process i(t) to be narrow-banded, Gaussian and


using the Longuet-Higgins distribution (as in Eq. 3.4) as the parent distribution of
wave amplitude, Turner and Baker (1988) have derived the following result for the
distribution of crest height of the most extreme wave

F(p) - exp[-NL(e)exp(-p 2 )] (3.9)

f(p ) - 2Np L(e) exp [ -p 2 -NL(c) exp(-p 2 )] (3.10)

where N is the expected total number of waves in the extreme storm. For any other
parent distribution of H the distribution of H, can be similarly derived. The period
of the maximum wave T, is usually taken as some deterministic function of T.
3-10

One of the main objections to this approach is that the method does not account for
the probability that the life-time extreme wave may occur in a sea-state different from
the most-extreme storm. An alternative approach would be to utilize the data on
observed maximum wave heights in each sea-state and derive the probability
distribution of the life-time maximum wave height from asymptotic considerations.
However, a truly "extreme" wave condition should be defined as the wave condition
which produces the highest response (forces/stresses) in the structure. This aspect is
discussed further in the context of extreme response modelling in Section 3.4.3.

3.2.4 Joint Occurrence of Wave, Wind and Current Conditions

Offshore structures are subjected to the simultaneous action of wind, waves and
currents and a joint probabilistic distribution of the three is required in determining
the set of extreme environmental parameters for the reliability analysis of these
structures. Uniform wind action is usually modelled in terms of the 'mean wind
velocity' with an averaging period of 10-minutes and measured at a reference level
of lOm above the sea-surface. The variation in this velocity with height is determined
using a power law function. More rapid fluctuations in wind velocity, called 'wind
gust', typically extending for a period of 1-60 seconds, can be modelled as a
zero-mean, stationary Gaussian process defined by a spectrum such as the Davenport
spectrum. Current is modelled in terms of the 'mean current velocity' at the sea
surface with an averaging period equal to the sea-state duration and a 'current profile'
to determine the variation with depth. More rapid fluctuations in current velocity
within this averaging period are usually not significant.

Since wind, wave and current vary with time and could act on the structures from
different directions, it is generally difficult to determine when the combined action of
the three will be a maximum. A more rigorous probabilistic representation would be
to model the significant wave height (and zero-crossing period), wind velocity and
current velocity and their respective directions as stochastic processes in the long-term.
Additionally, fluctuations in the sea-surface elevation and wind gust in the short-term
3-il
could be modelled as stochastic processes defined conditionally on the significant
wave height and mean wind velocity. Such a modelling would require time-dependent
reliability analysis methods based on out-crossing concepts. This method, although
possible, is not sufficiently well developed. An example of this approach can be seen
in Bjerager et al (1988) in relation to moored offshore structures.

However, for statically responding shallow-water jacket structures dominated by wave


loading it is reasonable to assume that maximum of the combined environmental
loading will occur when the wave loading is a maximum. Thus mean wind velocity
and current velocity can be defined conditional on the wave height. Based on a
20-year hindcast data of wind, wave and current for a Southern North Sea location
Haver and Winterstein (1990) have suggested modelling the mean wind velocity and
mean surface current conditional on the extreme significant wave height as lognormal
variables. Unfortunately, simultaneous observations of wind, wave and current
directions are not available to describe their joint distributions. Normally they are
modelled as independent of the wind, wave or current velocities defined by their
marginal densities. In response analysis wind, wave and current may all be assumed
to act from the same direction but with different directional probabilities.

3.3 METHODS OF RESPONSE ANALYSIS

The response analysis of offshore structures involves a complex set of calculations but
the procedure is well established as industry practice. A number of computer
programs are now available for automated response analysis and the procedure is well
described in a number of text-books. Only an outline of this procedure is presented
in this section to help discussion in later chapters. Two types of response analysis are
in common use for offshore structures, which are outlined below, and in principle
these can be used within a reliability analysis.
3-12

3.3.1 Deterministic Discrete Wave Static Analysis

In this method, the response of a structure is calculated corresponding to a single wave


characterized by the wave height and period. For this reason the response of the
structure has to be static in character, so that previous response history has no
influence on the response of the structure to the current wave. The first part of the
analysis consists of calculating the environmental loading on the structure expressed
in the form of distributed member loads or equivalent nodal loads. The second part
follows standard methods of stiffness matrix frame analysis.

A number of wave theories, such as Airy, Stokes, Cnoidal, Stream function theory
etc., are available, each of which has a range of applicable water depths, to determine
water particle velocities and accelerations at particular water depths, see for example
Sarpakaya and Issacson (1981). For slender tubular members Stokes fifth-order theory
is commonly used in moderate to deep water conditions.

The wave force, per unit length, on a cylindrical member is calculated using Morison
equation given by

dF - !pcdDlüItt+!pc mD 2 Ü (3.11)

where p is the density of sea-water, D is the diameter of the member, Cd is a 'drag


coefficient' and Cm is an 'inertia coefficient', ü is the velocity and ü the acceleration
of a water particle normal to the member induced by the wave. The force coefficients
Cd and Cm are empirical and are obtained from laboratory or full scale experiments.
These are observed to depend on factors such as relative surface roughness, Reynold's
number, Keulegan-Carpenter number etc. However, in most design applications
constant values of Cd and Cm are used for all members in the structure. Significant
uncertainties are associated with these values and in a reliability analysis Cd and Cm
should be modelled as random variables , as presented in Chapters 5 and 6.
3-13
The effect of steady current can be included in a simplest way by adding vectorially
the current velocity u to the wave induced particle velocity Up (U = Up + U) before
entering the Morison's equation above to calculate the combined hydrodynamic
loading. The wind force due to the uniform wind velocity u,,, on the exposed parts of
the structure is calculated using the drag term of the Morison's equation with
appropriate values of force coefficients. The combined environmental loading on each
of the members may also be expressed as equivalent nodal loads and the 'total
horizontal base shear' and 'overturning moment' about the mud-line computed. It is
clear that wave forces on various members vary as the wave is stepped through the
structure. The wave position (phase angle) which maximizes the base shear or
overturning moment is used in computing wave forces on members. Response
analysis is carried out using the stiffness method to obtain nodal deflections, member
internal forces or stresses.

3.3.2 Spectral Dynamic Analysis

In contrast to the previous approach a spectral analysis method determines the


short-term response of a structure in a random sea-state represented by the spectral
density function of the sea-surface elevation. Dynamic (inertial) behaviour of the
structure can also be included. The method involves the calculation of 'frequency
transfer function' relating the sea-surface elevation t(co) to the response parameter of
interest and then developing the full spectral density function of the response. The
response parameters of interest are usually nodal displacements, member internal
forces or hot-spot stresses in tubular joints. The calculation of the transfer functions
is done in several stages as outlined below. One of the fundamental requirements of
this approach is that the output parameter should be linearly related to the input
parameter at each stage of the analysis.

The first stage of the analysis involves calculation of 'wave force transfer function',
H(o,), relating the sea-surface elevation (w) to the wave force on a member (or
equivalent load at the end nodes) as a function of the wave frequency oi. Water
3-14
particle kinematics are obtained using Airy's linear wave theory which gives a linear
relation between particle kinematics and surface elevation. Wave force on a member
can be calculated using Morison's equation. However, the drag term in the Morison
equation has to be linearized to obtain a linear relation between wave force and wave
amplitude. A procedure called 'stochastic linearization' is commonly used in which
the linearized Morison equation is written as

(3.12)
dF _PCd')0ü4_PCnPU

where a is the root-mean-square (rms) value of particle velocity. The error involved
is expected to be small for cases in which wave loading is inertia dominated such as
for large diameter members well below the sea-surface and small amplitude waves.
However for slender members near the water-line and for extreme waves the wave
loading is drag dominated and the error involved in linearization could be significant.
Furthermore, since Airy theory is used wave loading above the still-water level and
effects such as wave slamming in the splash zone cannot be easily included.

The structural response transfer function, I-L(o), relating wave force on the structure
to a nodal displacement can be conveniently considered in two parts. The first
corresponds to the quasi-static response of the structure the calculation of which
largely involves operations on the structural stiffness matrix which are independent of
frequency and wave loading. This fact can be utilized to save considerable
computational time in the calculation of transfer functions and especially in a
reliability analysis when the basic variables involve mainly the parameters for
environmental description and calculation of wave loading, see Karadeniz (1989).

The second part corresponds to the dynamic response of the structure the transfer
function for which can be obtained using a modal analysis technique. The structure
is idealized as a lumped mass system with the structural mass lumped at the horizontal
framing levels with the deck mass at the top. The added mass of water surrounding
the structure which participates in the dynamic motion of the structure is also
considered. A linearized form of damping, consisting of structural, hydrodynamic and
3-15
foundation damping, is used. In the calculation of wave forces using the Morison's
equation it is preferable to consider the velocity of oscillation of the structure, t =u. -
ü1 and a in Eq.3.12. However, this requires an iterative solution since structural
velocity is not known in advance, see Maihotra and Penzien (1970) and also
Thoft-Christensen and Baker (1982, Appendix-B).

Finally the spectrum of the response, for instance axial stress SA in a member, is


G,3 (w) - H(o)G(w) (3.13)

The results of a spectral analysis are expressed in tenns of the moments of the
response spectrum. The nt) moment is computed using numerical integration as

m - (3.14)

Usually the zeroth, second and fourth spectral moments, corresponding to n =0,2 and
4, are of interest. A joint variation of K response components, for example nominal
stress components due to axial, in-plane and out-of-plane bend loading in a brace
member can be obtained in terms of their cross-spectral density functions as

G'I (w) - (w) H (w) G (0) i - 1,2,...,L ; j - 1,2,...,K (3.15)

in which * denotes a complex conjugate. The cross-spectral moments ni can be


obtained similarly to Eq.3.14. Finally the spectrum of hot-spot stress, which is a
linear combination of the K nominal stress components, can be obtained as

(3.16)
Gs(co ) - E E cjCjHsy,(w)H;y1(co)Gp(o))
s-I j-I

where C etc. are the 'stress concentration factors', see Chapter 4.


3-16

3.4 PROBABILISTIC DESCRIPTION OF STRUCTURAL RESPONSE

Since the environmental loading on an offshore structure is random the structural


response will also be random. The probabilistic description of response depends, to
some extent, on the probabilistic description used for environmental modelling and the
method of response analysis. Similar to wave condition modelling, the long-term
fluctuations of response can be considered in terms of a succession of responses to
short-term sea-states during which the response process can be assumed to be
stationaiy. The response quantities of interest here are the internal forces in members,
axial, in-plane and out-of-plane nominal stress components in brace members and the
hot-spot stresses in tubular joints. The short-term, long-term and extreme models for
these response quantities are considered in this section. For convenience all these
response quantities are denoted as S. The probabilistic modelling of hot-spot stress
ranges required for cumulative fatigue load estimation is considered in Section 3.5.

3.4.1 Short-term Response Modelling

A spectral density function for a response quantity, G5(o.), in a short-term sea-state


can be obtained from a spectral dynamic analysis of the structure as described in
Section 3.3.2. This response spectrum will however be conditional on the values of
the environmental parameters: H, T, 9, V and V, and the values of the other
uncertainty variables such as Cd, Cm etc., used in the analysis. The spectral
description of the response in frequency domain is characterized in terms of the
spectral moments m0, m2 and m4 . For a probabilistic description of the response in
amplitude domain certain additional characteristics are required. These can be derived
from the spectral moments as discussed below.

Since-the sea-surface elevation process T(t) is considered as stationary the response


process S(t) in the short-term can also be considered as stationary. Although the T(t)
process is adequately described as a Gaussian process, this assumption is not strictly
valid for the response. Because of the various non-linearities involved, e.g. in particle
3-17
kinematics, Morison's equation for hydrodynamic loading, soil-structure interaction
etc., the response of the structure becomes non-Gaussian. This has been observed in
some of the on-site measurements which show non-Gaussian response behaviour,
particularly for members in the splash zone where the wave force is drag dominated.
However, for most of the members below the splash zone where the wave force is
inertia dominated and since the response is derived using a linearized spectral analysis
procedure the computed response may be considered as an equivalent Gaussian
process. This provides a consistent and mathematically convenient model. The
following properties of the process may be obtained from the spectral moments, see
for example Newland (1975)

(3.17)
E[S 2] - a2 ^ji - m0

EE-a-m2 (3.18)

(3.19)
E[ 2] - a' - m4

where (t) is a first derivative process and (t) is a second derivative process of S(t).
R5 is the mean value of the process which can be obtained from a mean value static
response analysis of the structure (J.L8 = 0 when only wave forces are considered).
With t known the variance a or the standard deviation a of the process can be
determined from Eq.3.17. For a stationary process the mean value of (t) is always
zero and in the Gaussian case it can be shown that S(t) and (t) are uncorrelated.
Additional information about the process S(t) can be obtained from


T0 - .._. - 2it(m0/m2)112 (3.20)

T - 2.... - 2ir(m2/m4)''2 (3.21)


p Vp
3-18

T V0 ______
a - ._.. - - - ________ (3.22)
T0 v (m0m4)"2


e - ( l-a2)"2 (3.23)

where T0 is the expected time between successive up-crossings of the mean level
while T gives the expected time between successive peaks (maxima) of the process.
v0 and v, are therefore, respectively, the expected frequency of mean-crossings and
expected frequency of peaks. The ratio a is called a 'spectral irregularity factor' and
c as the 'spectral bandwidth' and gives an idea of frequency spread of the response
spectrum. When c—'O the spectrum is called 'narrow-banded' and for e —'l it is called
'broad-banded'. Although the sea-surface elevation process is approximately
narrow-banded, due to dynamic behaviour of the structure the high-frequency tail of
G,(co) spectrum for low-to-moderate sea-states is amplified resulting in a
broad-banded response spectrum, often with two or more peaks.

The probability distribution of peaks of a stress response process S(t) is of


considerable interest both for the prediction of extreme stress amplitude in a given
duration and for the estimation of cumulative fatigue loading. The peaks or local
maxima in a process S(t) occur whenever the derivative process S(t) has a
zero-crossing, i.e. changes from positive to negative. Considering a reduced process
Y(t)=[S(t)-p.]/, the probability density function for peaks is given by Rice's (1944)
well known result


f( y) - Lfy (y,o,z)IzIdz (3.24)

For a stationary, Gaussian process this simplifies to, see Madsen et al (1986)

f(y) £ exp[-(y/cIT)2 ] + ay4'(ay/c) exp(-y 2/2) (3r.25)


3-19

where 4e) is the standard normal distribution function. Note that this function
reduces to a Rayleigh density for a narrow-banded process (a1) and to a Gaussian
density for a broad-banded process (a=O).

3.4.2 Long-term Response Modelling

The short-term response process and all other short-term characteristics of the response
derived above are conditional on the values of I-I, T, e0, V and V 1, used in the
response analysis for a particular sea-state. In the long-term the environmental
parameters vary according to their joint-probability distribution, discussed in Section
3.2.4. Therefore, long-term response characteristics can be developed by carrying out
response analysis for a discrete number of representative sea-states and combining the
contributions of all sea-states duly accounting for the relative frequency of occurrence
of individual sea-states. Thus,for example, the long-term average frequency of peaks
can be obtained as

V E EV e P P( Tz 1s)PPYJ' 1:)P(8 ) (3.26)


O,j

where v8. is the short-term average frequency of response peaks for the j-th
sea-state with the wave approach angle 0 . obtained as in Eq.3.21. Note that the
probability of occurrence of T, V and V,, are defined conditionally on the value of
H1 as discussed in Section 3.2.4. The long-term distribution of response peaks can be
similarly obtained as

ft(s) -
• i (Y)_1 v p(Hs)p(T2It-I,)p(V)FI g )p(VJ1-I)p(0 j )
s P,, (3.27)

where the short-term density is additionally weighted by the average total number of
response peaks in that particular sea-state. The long-term density function is thus
available only in a numerical form. If necessary an analytical function such as a
Weibull model could be fitted to this distribution. A procedure for this is described
3-20

in Section 3.5.3 in the context of long-term distribution of hot-spot stress ranges. The
long-term distribution of response peaks is also of interest in extreme response
modelling discussed next.

3.4.3 Extreme Response Modelling

The probabilistic modelling of extreme response depends to some extent on the


environmental modelling and the method of response analysis used. In general the
models could be categorized as short-term extreme and long-term extreme models.

The simplest approach is to assume that the life-time extreme response of a structure
occurs when the highest wave in the life-time extreme sea-state passes through the
structure. In this case the deterministic discrete wave static analysis could be used for
response estimation. The probability distribution of extreme response then follows
from the probability distribution of the most-maximum wave height (and the
associated wave Perioci) defined conditionally on the parameters of the extreme
sea-state, (Hs)E and (TZ)E. This approach is consistent with conventional design
practice and has been used in some of the published studies on reliability of offshore
structures, see for instance Turner and Baker (1988). This approach has also been
used in the development of RASOS software, see Goyet et al. (1991) and will be
discussed further in Chapter 6. The use of deterministic static analysis for response
estimation gives considerable computational advantage in a reliability assessment.

The above approach seems to be reasonably adequate for statically responding


shallow-water jacket structures. However, for deep water structures where dynamic
response is predominant a spectral dynamic response analysis is invariably required.
Furthermore, since the response process is usually broad-banded the maximum
response amplitude may not correspond with the maximum wave in the sea-state. In
such a case the structural response may be defined conditional on the parameters of
the extreme sea-state assuming, however, that extreme response occurs in the extreme
sea-state. The distribution of all response peaks then follows from Eq.3.25. The
3-21
disiribution of the largest maximum response amplitude in a sea-state duration At can
be derived using order statistics. For a broad-banded, stationary, Gaussian process
Cartwright and Languet-Higgins have shown this to follow an extreme type-I
distribution given by, see Madsen et al (1986)

FE (yE) - exp[-Nexp(-y2/2)]

- (2 logN)1 "2 + [yI(2 logN)"2]


(3.28)
- (it2/6)(lI2logN)

N - avAt

where =O.577 is the Euler's constant. It is to be noted that this distribution for the
extreme response applies to individual stress response components in a member and
also to the hot-spot stress which is a linear combination of these stress components.
However, for the von Mises equivalent stress process which is obtained as a non-linear
combination of stress components Gaussian assumption cannot be used and this
distribution does not apply. Madsen (1986) has derived the extreme distribution of
von Mises stress using an out-crossing approach.

The main drawback of both of these short-term extreme models is that they do not
account for the possibility that extreme response of the structure may occur in a
sea-state other than the life-time extreme sea-state. When dynamic behaviour of the
structure is significant, depending on the relative magnitudes of wave frequencies and
natural frequency of the structure the extreme response of the structure may occur in
a sea-state less severe than the life-time extreme sea-state. Furthermore, since the
environment affects different responses in different ways it is not possible to
determine in advance which set of environmental conditions will produce the largest
maximum response. In such a case the most rigorous approach would be to develop
the complete long-term distribution of response peaks using the structure's response
to all likely sea-states as given by Eq.3.27. The distribution of the long-term extreme
response amplitude can then be obtained as
3-22

(3.29)
FE(sE) -
Iii -[1

where T is the service-life of the structure (in seconds) and vL is the long-term
average frequency of peaks given by Eq.3.26. In this case both long-term and extreme
distributions will, however, be available only in a numerical form.

3.5 PROBABILISTIC MODELLING OF STRESS CYCLES FOR FATIGUE

Fatigue damage in offshore structures occurs at the welded tubular joints mainly as
a result of the cyclic action of ocean waves. Fatigue is a cumulative phenomenon
characterized in tubular joints by the initiation and propagation of fatigue cracks from
weld defects located in regions of stress concentration. An S-N curve approach or a
fracture mechanics approach is used for quantifying fatigue damage in tubular joints
and these methods are discussed in detail in Chapter 4. For the type of steels used in
offshore construction it has been observed that the stress cycle interaction effects are
usually negligible and in such cases the cumulative effect of fatigue loading due to
long-term wave conditions can be conveniently calculated independent of fatigue
resistance of the joint. In both the methods of fatigue damage assessment mentioned
above this can be represented in terms of a fatigue loading function given by

n(T) (3.30)
IVL (T) - L

where S1 is the stress-range magnitude for the th stress cycle, m is either the inverse
slope of the S-N curve or the exponent in the Paris law for crack growth, see
Chapter 4 for details. The summation is taken over all stress cycles n(T) likely to
occur during a service exposure of duration T.

A more direct approach for calculating cumulative loading is to simulate the stress
time-history and count the individual stress cycles and stress-ranges. Stress cycles can
3-23

be easily identified in a narrow-banded stress time-history but for wide-banded stress


processes, such as those of hot,tress of offshore tubular joints, individual stress cycles
cannot be easily distinguished. In such cases a cycle-counting convention has to be
adopted. Dowling (1972) has reviewed a number of cycle-counting methods of which
the "rain-flow cycle counting method" is now widely accepted to give best results for
fatigue damage predictions. However, a cycle-counting approach is computationally
very expensive and for practical applications a much simpler approach is desirable.

In Eq.3.30, although the individual stress-ranges are random, for large numbers of
cycles n(T) the variance of the sum, being inversely proportional to in(T) , is very
small and the summation term can be adequately represented by its expected value as

n(T)
'VL (T ) E[E Sam ] - E[n(T)] E[Sm] (3.31)
i-I

The expected number of stress cycles E[n(T)] and the mth expected value of
stress-range E[S m] can be easily computed if the probability density function of
stress-ranges is known. Thus most of the simpler and indirect approaches for
computing cumulative fatigue loading attempt to develop a probability density function
for stress-ranges. Some of the available methods are reviewed in Section 3.5.1 and
a new method suitable for fracture mechanics based fatigue reliability assessments is
developed under Section 3.4.2, 3.5.3 and 3.5.4.

3.5.1 Review of Available Methods

A number of methods have been reported in the literature for efficient calculation of
cumulative fatigue loading which do not need expensive time-history simulation and
rain-flow cycle counting. These methods can be broadly grouped into two categories.
The first of these requires only a deterministic method for response analysis while the
second group of models are based on a spectral method of response analysis. These
are discussed below.
3-24

3.5.1.1 Methods Based on Deterministic Response Analysis

One of the widely used methods in this category is the method based on "wave height
exceedance diagram" suggested by most design codes for fatigue analysis. The
long-term distribution of wave heights is expressed as a plot of wave height versus
number of exceedances. The wave height axis is discretised into about 6-10 blocks
and global structural analysis is carried out for the mid-height of each block to
determine hot-spot stress-range. The number of stress cycles is then assumed to be
equal to the number of waves in the corresponding wave height block.

In the method suggested by Nolte and Hansford (1976) the long-term variation of
individual wave heights is assumed to follow a Weibull distribution, the parameters
of which can be established from the oceanographic data. The hot-spot stress in a
joint is then assumed to be linearly related to wave height so that the long-term
distribution of individual stress-ranges can be derived from the long-term distribution
of wave height. This method gives a closed-form expression for cumulative loading.

In the method suggested by Marshall and Luyties (1982) and adopted by the API code
(1989) for the Gulf of Mexico wave climate, the variation of stress-ranges is modelled
by a 3-parameter Weibull distribution the parameters of which are obtained from
detailed calibration studies on typical shallow-water (<zl2Om) structures in the Gulf of
Mexico. The user has to obtain the stress-range corresponding to the 100-year return
wave which provides the required scaling and the expected number of stress cycles
in the design life of the structure. The shape parameter, which expresses the severity
of fatigue loading with respect to the 'peak' loading, is obtained from calibration
curves depending on water depth, type of structures and position of the joint in the
structure etc.

Afl the methods discussed above are simple to use but are suitable for shallow water
structures where the response is linearly related to wave height. However, for most
structures in moderate water depths where dynamic effects are significant a spectral
response analysis is required.
3-25
3.5.1.2 Methods Based on Spectral Response Analysis

If the stress response process is stationaiy, Gaussian and narrow-banded it can be


shown that stress peak amplitudes and hence stress-ranges follow a Rayleigh
distribution. For this case cumulative fatigue loading function can be obtained as

'VL,NB() - E[n(t)] E[Sm] - VO t(2O bI) m JT(!!.+l) (3.32)

where a is the r.m.s value of the stress process, rr is the Gamma function and t is
the duration of the sea-state in seconds.

Because of this simplicity Rayleigh distribution is widely used in design practice for
modelling stress-ranges in tubular joints. However, when the stress process is
wide-banded Wirsching and Light (1980) found that Rayleigh assumption over-predicts
fatigue damage by about 15-20%.

In recent years a number of stress-range distribution models have been reported in the
literature for wide-banded processes. The basis of all these models is empirical. The
approach used is to simulate time-histories of processes with varying band-widths,
compute the cumulative fatigue loading using rain-flow counting and fit an empirical
model to simulation results. Expressions for fatigue loading function using some of
the reported models are presented below. These models are used for comparison with
the proposed new model in Section 3.5.2.3. All these models are applicable for stress
cycles in one short-term sea-state of duration t seconds. For computation of
cumulative fatigue loading function for the entire service life T of the structure
contributions from all sea-states has to be summed up duly accounting for the relative
frequency of occurrence of each sea-state.
3-26
Wirsching and Light (1980)

VL( t ) - X'WL.NB(t)
- a+(1_a)(1_c)b
(3.33)
a 0.926 - O.033m
b - 1.587m - 2.323

Hancock et at (1986)

VL( t ) - v tAmr(.^1)
B
(3.34)
A - (2-e2)aJ
B - 2-c2

Choudhury and Dover (1985)

I c+2
VL( t ) - Vpt [ 2am L 2 r m+1
2
+ 1a(m+2)]
(3.35)
j
with 1 - 0.75

Kam and Dover (1988)


Same as Eq.3.35 but with 'y given by

- ..[1+e,f(y)]
(3.36)
- + 0.3012cz + O.4916ct 2 + O.9181cz 3 - 2.3534a4
2
- 3.3307a5 + 15.6524a6 - 10.7846a7]

Zhao and Baker (1990)

L(t) - v t (2a) m [ya m r( ..+ 1) +

a - (8_7a)_h1'

b - 1.1 for a^0.9 (3.37)

- 1.1 +9(a-0.9) for a>0.9

7 - (1-a) / [1-ga F(!+1)]


3-27
3.5.2 Proposed Model for Short-term Distribution of Stress-Ranges:

As mentioned earlier the hot-spot stress in a tubular joint is generally observed to be


broad-banded. In order to avoid expensive time-history simulation and rain-flow cycle
counting a simplified model for obtaining the cumulative load-effect in the short-term
IS recurct
in a given sea-stateL The proposed method requires the hot-spot stress spectrum to be
obtained using a spectral response analysis of the structure. A probability density
function for modelling the short-term stress-ranges is first developed in
Section 3.5.2.1. An expression for cumulative fatigue loading function based on this
density function is derived in Section 3.5.2.2. The proposed model is compared with
simulation results and other available models in Section 3.5.2.3.

3.5.2.1 Probability Density Function for Stress-Ranges

The probability density function for stress-ranges in a broad-banded stress process is


obtained based on the theoretical density function for peaks of a broad-banded,
Gaussian process, described earlier in Section 3.4.1, and the peak-counting method for
defming stress cycles. In general the following assumptions are made:

(a) The hot-spot stress process in a tubular joint is assumed to be Gaussian. This
assumption is strictly not valid because of the various non-linearities affecting
wave loading and structural response. However, for reasons mentioned in
Section 3.4.1 the hot-spot stress process may be considered as
equivalent-Gaussian. The density function of stress peaks (or local maxima)
then follows from Eq.3.25.
(b) The stress process is assumed to be symmetric, i.e. the number of local
maxima below the mean level are assumed to be equal to the number of local
minima above the mean level. This assumption is fairly reasonable for linear
systems, regular waves and a sufficiently long time-history.
(c) The rate of cycle occurrence is taken to be the rate of occurrence of positive
peaks in the stress process. A local maxima above the mean level is paired
with a local minima below the mean of approximately the same magnitude to
3-28

Prob. Den.ity Function

0.8

06

0.4

0.2

0_O n • II
I I
II I
I
I I I

0.0 0.4 05 12 1.8 2.0 2.4 2.8 3.2 3.5 4.0


Normali.ed Streis

Fig.3.1: Probability density functions of all peaks and positive peaks

form a stress cycle regardless of the position of these maxima and minima in
the time-history. This means that all cycle-interaction effects are ignored.
Cycle interaction effects are observed to be relatively unimportant for
structural steels. Furthermore, if the time-series is sufficiently long it is
possible to fmd for every local maxima above the mean a local minima of the
same magnitude below the mean level.

With the above assumptions the probability density function for hot-spot stress-ranges
can be obtained from the density function of positive peaks by taking stress-range to
be twice the amplitude of a positive peak. Now the density function of positive peaks
can be obtained from the density of all local maxima (peaks) as follows. Integrating
Eq.3.25 from - 00 to 0 gives the proportion of negative maxima in a time-history as
(1-a)12. Therefore the proportion of positive maxima will be (1+a)/2 and hence the
probability density function of normalised positive peak amplitude can be given as

f;(y) ..f,,(Y)/(1;cL)
fory>0 (3.38)
-o fory<0
3-29
where f.,,(y) is the density function of all local maxima given by Eq, 3.25. The density
functions f.7(") and f(') are shown in Fig.3.1. Now the density function of
stress-ranges can be obtained as

f(s) - f;( y ) - .__.


2a f( y) - ( l+a)'W (3.39)
dS

3.5.2.2 Expression for Cumulative Fatigue Loading

The cumulative fatigue loading function is given, as before, by Eq.3.31. The expected
number of stress cycles in the sea-state duration t seconds is taken as the expected
number of positive peaks in the time-history. The expected total number of peaks is
given by v,t while the proportion of local maxima below the mean level (negative
peaks) can be easily shown to be (l-a)/2. Therefore the number of positive peaks and
hence the expected number of stress cycles is given by

2 Jv p t
+ (3.40)
E[n(t)] -
'L.

The mth expected value of stress-range is derived from the probability density function
in Eq.3.39 as follows

E[Sm] fs mf(S)dS - f(2ay)mf;(y)cy - (2a)mE[Ym]' (3.41)


0 0 [(l+a)/2]

It is thus Convenient to use the density function of all local maxima ç(y) from Eq.3.25
and compute the partial expectation of the normalised variable Y as

= I_
E[Y fyml_e exp[-/c)2] + aY(aY/c)exp(_Y2/2)} (3.42)
-
01-
3-30
Unfortunately an exact analytical solution of this integral is not available. Although
numerical integration can be used, it is preferable to have a closed-form analytical
expression for E[S m } for use in incremental fracture mechanics fatigue crack growth
calculations and this is attempted below. The following derivation was inspired by
an earlier work of Kam and Dover (1988).

In the first step Eq.3.42 can be expressed as

E[Y] - ( / ) m1 C"'2 F( m+l ) + a fy m^1 exp( _y 2, 2)(ay,c)dy (3.43)

The integral in the second term does not have a simple analytical solution but can be
closely approximated by

afy m+1 exp(-y 2/2)(cty/e)dy ()mcx b(ay/e) ( m+2 (3.44)


2

The difference between this approximation and exact numerical integration is observed
to be less than 5% for most cases. Substituting this into Eq.3.43 leads to

(/)m1 Cm*2 (m+l) + (%/)ma(ay/c)f(m+2)


E[Ym] - (3.45)
2

This equation still contains the normal distribution function () which acts as a
weighting factor for the second term in Eq.3.45 and is found to be a function of the
normalized stress-range y. This has arisen because of the particular approximation
used in Eq.3.44. If Eq.3.43 is evaluated exactly the y-term would have been
integrated out and the resulting expression would be independent of y. Therefore for
using Eq.3.45 the weighting function should be evaluated at a suitable value of y. The
choice should be such that the resulting error in evaluating E[Ym]° from Eq.3.45
should be the least when compared to exact numerical integration of Eq.3.42.

A number of options for evaluating 4(') have been compared in an earlier paper by
Shetty and Baker (1990). It is found that Eq.3 45 can be evaluated exactly if(1(') is
calculated at the "equivalent normalised stress-range" y defined as
3-31


Yeq (E[ym1)l/m (3.46)

Based on Eq.3.45 1(') can be evaluated at as

ay - Ja [()m I e2 r(m+l) + ___ ___


£

(3.47)
This results in another normal distribution function 2() and Eq.3.47 has to be
solved by iterating on ('). This hardly presents any problem and experience shows
that convergence can be obtained within 3-4 iterations. However, if a direct
evaluation of c1() is desired in practical applications, it is found that evaluating 4()
at the equivalent normalised stress-range of a narrow-banded process and then
substituting in Eq.3.47 gives a very good approximation for 1('), i.e.

(3.48)
NB[] - { WT m r(m;2)]1/m}

In an iterative procedure 'NB() from above can be used as a starting value for I).

The final expression for VL(t) can now be written as

(l;a)Vt(2af)
____ m[m+2
_____ m+l )+aT( m+2)(aY)] 1
'I'L(t) -
2%/ 2 2 li-a

(3.49)
with b(') obtained from Eq.3.47. It can be observed that for the case of a pure
narrow-banded process with a=1 this expression reduces to Eq.3.32 obtained based on
a Rayleigh density function.

3-32
3.5.2.3 Comparison with Previous Models

Predictions of the proposed model are now compared with the previous short-term
models presented in Section 3.5.1.2. All these models can be compared on a
consistent basis in terms of the parameter ). giving the ratio of cumulative loading
predicted by a model to that given by a narrow-banded process as given in Eq.3.33.
Wirsching (1980) has presented 11 stress response spectra typical of offshore
structures subjected to fatigue loading due to waves. These spectra extend from near
narrow-band to near broad-band and offer a good test to compare the predictive
capability of any model. These spectra have been widely used for comparisons in
previous studies and simulation results with rain-flow counting are available for these
spectra. For this reason the same 11 stress spectra are chosen here for comparisons.
The simulation results are obtained from Thao (1989). The 11 stress spectra with the
associated response parameters are reproduced in Table 3.1.

Table 3.1: Sea-state and spectral characteristics proposed by Wirsching.

Significant Zero-crosng. Irregularity Bandwidth


Sea state wave height period factor factor
__________ H (in) T (sec) a e
1 16.01 17.30 0.512 0.859
2 14.48 16.50 0.539 0.842
3 12.96 15.80 0.565 0.825
4 11.43 14.70 0.610 0.793
5 9.90 13.60 0.657 0.754
6 8.38 12.70 0.699 0.715
7 6.86 11.60 0.752 0.659
8 5.33 10.30 0.815 0.580
9 3.81 9.10 0.869 0.496
10 2.28 7.70 0.922 0.388
11 0.76 4.40 0.988 0.152

Plots of ? vs m given by different models and from simulation with rain-flow


counting are shown in Fig.3.2 for different response spectra. These plots show that
predictions of the proposed model are in reasonable agreement with previous models,
being slightly on the conservative side.

3-33

Damage Correction Factor


1.4
1.3 Expon.nt m.3

1.2
1.1 N

1 : ------
0.9 Z

0.1
0.6
0.6
0.4
0.3
I I I
0.2
.606 .633 .669 .603 .661 .692 .148 .808 .862 .917 .986
Irregularity Factor

Fig.3.2: Comaparison of fatigue damage ratio from different models

Prob. D.nsity Function


1.4

1.2

0.8

0.6

0.4

0.2

0
O.0 0.4 0.8 t2 tO 2.0 2.4 2.8 3.2 3.6 4.0
NormaIis.d Btr.s Rang.

Fig.3.3: Comparison of stress-range density functions from different models


3-34

The predictions of all the models are in good agreement with simulation results, being
conservative in general. The rain-flow correction suggested by Wirsching and Light
(1980) reduces the conservatism of the narrow-band assumption considerably but the
agreement with simulation results is not good in all cases. The Zhao and Baker
(1990) model gives by far the best agreement with simulation results.

The Choudhury and Dover (1985) model and Kam and Dover (1988) model are
similar to the present model as they both use the same probability density function of
stress peaks. But the underlying assumptions involved and the method of counting
cycles to obtain the probability density function of stress-ranges have not been noted
before. Moreover, additional approximations have been introduced in deriving the
expressions for cumulative loading function. It can be easily shown that the
Choudhury and Dover model is equivalent to using a constant value of 0.75 for
c1(ay/e) in Eq.3.45 while the Kam and Dover model is equivalent to evaluating (I)(')
at the expected value of y. Thus the b(') function in both these models is
independent of m resulting in under-estimation of E[Sm] compared to exact numerical
integration. Interestingly, this under-estimation results in a better agreement of the
Kam and Dover model with simulation results. The present model, although close to
exact numerical integration, is conservative when compared to simulation results and
this is perhaps to be expected of any model based on peak-counting.

Next, the probability density function of stress-ranges predicted by different models


are compared with simulation results in Fig.3.3 for sea-state 4. It can be seen that the
density function based on the distribution of positive peaks and the Rayleigh model
based on narrow-band assumption give very poor comparison with results from
rain-flow counting. Because of this, the cumulative damage predicted by the present
model and the Kam and Dover model are likely to give less favourable comparison
with simulation results when partial expectations are used, as in the case of fatigue
threshold effects or corrosion fatigue crack growth ana1)'is. In such cases the
empirical model of Zhao and Baker should be preferred as it gives the best
comparison with the stress-range density function from rain-flow counting.
3-35

3.5.3 Proposed Model for Long-term Distribution of Stress-Ranges

Fatigue loading caused by a single sea-state can be calculated using the short-term
stress-range model presented in the previous Section. If the long-term wave climate
is represented by M discrete sea-states then the cumulative fatigue loading over a
service exposure of T seconds can be calculated by a weighted summation of load
contributions from individual sea-states as

M
'I'L(T) - E[Sm]v,PT (3.50)
i-i

where P3 is the relative frequency of occurrence of the th sea-state.

However, when the number of sea-states considered is large (>20) and when threshold
effects are important the computation of partial expectations during incremental
crack-growth calculations (see next Section) becomes computationally very expensive.
In such cases it is desirable to have a single, analytical distribution to model
stress-ranges in the long-term. Because of non-stationarity and the choice of arbitrary
number of sea-states it is not possible to derive a universal long-term model which
will be applicable in all cases. The approach used here is to assume a general
distribution function and obtain the parameters of the function by a fit to the
short-term stress-range data using one of the distribution fitting techniques.

From a careful observation of the form of short-term density functions for varying
band-widths a mixed-Weibull distribution is chosen to model the long-term variation
of stress-ranges. The density function is given by

D (S)D1
fi(s) - W(!)91ex [(S)B] + ( 1-W)_ exp[()"] (3.51)
A
3-36

where A, B, C, D and W are the parameters of the density function to be obtained


from fitting this model to short-term data.

The long-term average frequency of stress-cycles can be easily obtained from the
short-term data as

M
1+cx
CD1 -
E
i-I
2
)vP (3.52)

Based on the mixed-Weibull density function the mth expectation of stress-range and
hence the cumulative fatigue loading function for a service exposure of T seconds can
be obtained as

E l [S m ] - WA m r(1+.?.!.) + (1_W)CmT(1+!!) (3.53)


B D


VL(T) - E l [S m ] CD I T (3.54)

In the following two model fitting techniques are used to obtain the unknown
parameters A, B, C, D and W.

3.5.3.1 Method of Moments

This method is commonly used for fitting distribution models to statistical data. In
the present the 5 unknown parameters of the Weibull model are determined by
matching the first 5 moments of the long-term Weibull model with those computed
by a weighted summation of moments for the M sea-states based on the short-term
density function model. Thus

EE[Sm]v,P - E l Es m lw l m-1,2,...,5 (3.55)


j-1
3-37
where E(') is the mth expectation calculated using the short-term model Eq.3.41 and
E'(') is computed from the Weibull density using Eq.3.53.

The resulting set of 5 non-linear simultaneous equations can be solved using the
Newton-Raphson iterative technique, see Press et al (1986). This technique has a
rapid rate of convergence but when the M sea-state spectra span a wide range of
band-widths it is likely that convergence may not be obtained within close tolerance
limits. In such cases it is preferable to determine the Weibull parameters which
minimize the squared error. Thus

5
Minimize E [z3(m)-z,(m)]2
rn-i
(3.56)

in which ; corresponds to the L.H.S and ; corresponds to the R.HS of Eq.3.55.


This non-linear minimization can be carried out using the Levenberg-Marquardt
method, see Press et al (1986).

3.53.2 Method of Chi-square Minimization

In this method a stress cycle histogram is first developed by a weighted summation


of stress cycles from all the sea-states using the short-term density function model.
A histogram with variable bin (interval) sizes is used to accurately represent stress
cycle density in all sea-states and at the same time keep the total number of bins to
a minimum. If there are a total of L bins, the total number of stress cycles N in an
arbitraiy time interval r, and the number of stress cycles nj for the ith bin can be
obtained using the short-term data as

M
N - __ (3.57)
h_il
2 IJvjPJt
3-38

M

(3.58)
- Ef(y)dy1vPt
j-1

where y1 = s1 /2c and dy1 = ds1 /2cY in which s1 is the stress-range corresponding to
JUl stress spectrum.
the th bin, ds1 is the bin size and is the r.m.s value of the

Using the long-term Weibull model the corresponding number of stress cycles for the
th
bin is

(3.59)
Np1 - Nf(s1)ds1

in which f (s) is obtained from Eq.3.51.

Treating the histogram obtained form the short-term density function model as a
sample of data a chi-square function can be constructed as

L
x2- (3.60)
_______
ii [(Np,)'2]

where the argument of the summation represents the normalised squared error in
fitting the model to the data- set for the ith bin. Thus if the Weibull model parameters
are chosen so as to minimize the x 2 value the resulting long-term model will give a
good fit to the short-term data.

The values of the 5 unknown model parameters can be determined by minimizing the
X2-function using the Levenberg-Marquardt non-linear minimization method, see Press
et al (1986). The vector of derivatives and the curvature matrix, required by the
minimization algorithm, are calculated by differentiating the X2-funCtiOn

- I(n-Np) + (fl1_NP)2lap
- (3.61)
Np1
•-; -' L 2Np
3-39

(ns_Npj21 ap, ap,


-
H: hJk
1a2x2
2 aak aaf
-
t ^ Np j•;;
(3.62)

in which a: j = l,2,...5 correspond to the Weibull parameters.

At the solution point it can be shown that Xm 2 follows a chi-square distribution for
(L-5) degrees of freedom. Thus the "goodness-of-fit" of the long-term model with the
short-term data can be quantitatively determined.

In determining the long-term Weibull model parameters as above it is assumed that


the stress-range histogram obtained using the short-term model represents the "true"
long-term variation of stress-ranges. Thus the model parameters determined will be
"unique" for the considered short-term stress histories. However, when the short-term
time-histories are random, as a result of random sea-state parameters and the various
uncertainties involved in determining response transfer functions and additionally
uncertainties in the short-term density function model itself (compared to rain-flow
analysis), the computed Weibull parameters can, at best, be regarded as "estimators"
of the "true" long-term model of stress-ranges. In such a case it is important to
estimate the covariance matrix of the estimated parameters and if possible their
joint-probability distribution. The chi-square minimization method can be used for
this purpose. It turns out that the joint-probability distribution function of the 5
long-term Weibull parameters can be used to represent the uncertainties involved in
determining fatigue loading in offshore tubular joints. An application of this method
is discussed in Chapter 5.

Compared to the method of moments the method of chi-square minimization gives a


very good fit to the short-term data as the fit is established for each bin in the
histogram. However, the method of moments is computationally much faster. In fact
a combination of the two methods, in which the method of moments is first used to
obtain reasonably good estimates of the parameters which are then improved by the
chi-square minimization method, gives an efficient method for determining the
Weibull parameters. This strategy has been used in the computer program "WEIBlJL"

3-40

developed by the author which has also been implemented in the program module
RASOS-H, see Birades and Shetty (1991). A comparison of the density function
obtained by the weighted short-term model and the fitted long-term Weibull model for
the 11 stress spectra of Table 3.1 is shown in Fig.3.4.

0.04

= 0.03

0
002

2 0.01
0.

0.00
c p. - .- • - , . r, .-
- • 0 C CO - - C . CO
- -o r • • C
-

Stress Range

Fig.3.4: Comparison of the long-term model with the short-term histogram.

It is to be noted that once the long-term Weibull density function is developed, the
short-term density function model is not required in the computation of the long-term
mth expectation of stress-range and the long-term cumulative loading function.
3-41

3.5.4 Incorporation of Stress Threshold Effects

In fatigue damage assessment using an S-N approach a threshold stress-range or an


endurance limit, s. is sometimes specified. All stress cycles with range s ^ s. are
regarded as not to cause any fatigue damage. Similarly in a fracture mechanics
approach a threshold stress intensity factor range, zK..b is often specified. For a given
crack depth a the corresponding threshold stress-range can be determined as, see
Chapter 4 for details

s h(a) - tiKlh(a)/Q(a)(lca)"2 (3.63)

where Q(a) is a factor dependent on crack geometry and loading mode. Thus crack
will not propagate for any stress cycle with range s ^ s,(a). Note that sth(a) here is
a function of crack depth and will change as crack propagates.

In computing the cumulative fatigue loading, therefore, only damaging stress cycles
with range s> s. should be considered. This requires determining the number of
damaging stress cycles and mth expectation of stress-range using the probability
density function of damaging stress cycles only. However, following the approach
used in Section 3.5.2.1 this density function can be expressed in terms of the density
function of all stress ranges. The cumulative loading function due to damaging stress
cycles can be obtained as, at a given crack depth

L( t ) - E [ n (t)] d EES m ] d - E[n(t)] E[S m ]° (3.64)

where the suffix 'd' denotes damaging stress cycles only. The partial expectation can
now be calculated using the original density function of all stress-ranges as

J m f () ds
E[Sm] (3.65)
-
3-42
When the short-term density function model is used this partial expectation can be
calculated as

E[S m ] - (2a)mE[Ym]/[(1+cL)/2] (3.66)

where y = s.jJ2a and for the normalised variable

E[Y m ];° - fymf(y)y

(3.67)
- Y m{ £_exp[_(y/c)2] # aY(aY/c)exP(_Y2/2)}

Following a similar approach as that in Section 3.5.2.2 integration of this leads to

+2 (m+1 ;
E [ Y tm r - ()m1 C
Y) + ( /T) b '(cty/c) r(m;2 ; Y1)
f2
(3.68)
r(k; u) - fx exp(-x) dx

is the incomplete gamma function, see Abramowitz and Stegun (1965). The function
cb'(') is now

ayeq ) - Ja [() m l Cm42 f(m+l


2 + () m ( eq ) m+2 iim}
'Yth)"
C 2

(3.69)
When the long-term wave climate is represented by a large number of sea-states, this
partial expectation is to be computed separately for each sea-state and for each crack
increment in a fracture mechanics approach. This could become computationally very
expensive especially in an iterative reliability analysis. In such cases it is more
feasible to use the long-term density function model for which the partial expectation
in Eq.3.65 can be computed based on Eq.3.51 as
3-43

E[S m ]° - 1S m SW B (! ) 1 CXpI_(-) 8 l + (
l-W) D (5)D1exp[_(S)D]}d
J AA

m+B " (m+D


- WA tm r( SgJ + ( l-W ) C r D ;
B (3.70)

For the general case of a bi-linear S-N curve or for the multi-segment da/dn curve
used for corrosion-fatigue crack growth assessments, the above approach of partial
expectations can be easily extended. The partial expectation and hence fatigue loading
sum due to stress cycles with ranges falling under different segments of the curve are
computed separately and added together. For a bi-linear S-N curve, for example.
using the long-term Weibull model E[Sm] can be computed as

+ E[S 1P ] 2 - E[S"]° + ( 3.71)


E[S m ] -

The individual partial expectations can be calculated similar to Eq.3.70.

The proposed procedure for calculating cumulative fatigue loading for offshore tubular
joints is thus rigorous, efficient and general for use in most practical cases of interest.
Chapter 4
MECHANICAL MODELS FOR CAPACITY OF TUBULAR JOINTS

4.1 GENERAL

The models presented in the previous chapter help to determine response of an


offshore structure in terms of member end forces or stresses. In order to formulate
safety margin equations for individual components it is additionally necessary to
determine the resistance of members and joints. Mechanical models to determine the
capacity of tubular joints are presented in this chapter. Strength models for primary
members are briefly discussed in chapter 6.

Following the introductory discussion on behaviour of primary members and joints in


chapter 1, it is clear that welded tubular joints of offshore structures can fail in one
of the following three primary modes of failure:

1. Plastic collapse under static loading


2. Fatigue failure under cyclic loading
3. Fracture failure under static loading

Fracture failure or plastic collapse of a joint can occur either with an initial defect of
a particular size or following some amount of fatigue deterioration. For reliability
analysis, three limit-states can be defined corresponding to these three modes of joint
failure. Resistance models for the three modes are presented in this chapter while
reliability analysis for these limit-states are presented in chapter 5.

It is to be noted that the above three aspects of tubular joint behaviour are currently
the topics of intense research. A review of the available literature shows that the
research in these three areas has advanced by unequal levels often ignoring the
interaction of each other and as a result a fully consistent model for assessment of
tubular joints is not easily available. Based on this review and further development,
an attempt is made in this chapter to propose a fully consistent, comprehensive and
4-2
yet reasonably simple mechanical model for strength of tubular joints. Emphasis is
given to the use of simplified parametric formulations which are suitable for use in
reliability analysis of jacket structures.

The model to determine the resistance against plastic collapse of an intact tubular joint
is based on the mean value formulae developed by the Dept. of Energy (1989b,
0TH 308). Collapse behaviour under individual modes of loading are discussed and
a simplified model for determination of the strength of a joint damaged by fatigue is
suggested in section 4.2.

Formulations available for determination of stress distribution, both in the through


thickness direction and along the welded intersection, of a tubular joint are
summarized in section 4.3. Idealizations for residual stresses and weld notch stresses
are also given which are necessary for use in a fracture mechanics approach.

A fracture mechanics fatigue crack propagation model has been developed in


section 4.4 to estimate crack growth through the thickness and at the surface of a
tubular joint. The proposed model uses simplified formulations for stress intensity
factors but explicitly accounts for the effect of weld geometry, residual stresses, crack
coalescence, load shedding, variable amplitude loading and fatigue threshold. The
proposed model is compared with experimental results of compliance function
variation, crack depth development and crack shape development and a reasonably
good agreement is observed.

The available methods of elastic and elasto-plastic fracture are reviewed in section 4.5.
A methodology for fracture assessment of tubular joints is developed based on the
CEGB R6 approach. The proposed method is compared with the few available test
results on as-welded and stress-relieved joints. It should be noted that explicit fracture
assessments are not currently used as a part ofthe design process for jacket structures
and as such no guidelines are available especially for tubular joints.
4-3

4.2 PLASTIC COLLAPSE OF TUBULAR JOINTS

Tubular joints are subjected to very high transient loads when an offshore structure
experiences a severe storm. The capacity of a joint to withstand these high static
loads is one of the important criterion considered in the design of jacket structures and
most codes offer guidelines for determining "design static strength" of tubular joints.

Knowledge of the ultimate behaviour of tubular joints has traditionally been obtained
from static load tests on scaled models. Attempts have been made in recent years for
using non-linear finite element methods. The capacity of a joint and the mechanism
of final failure is seen to be strongly influenced by the geometry of the brace and
chord members forming the joint, the presence of neighbouring braces and the mode
of loading applied to the joint through the brace member(s).

Over the years a number of empirical formulae for tubular joint capacities have been
presented in the literature. Reber (1972) and Pan et al (1976) are indicative of early
attempts to develop formulae for joint strength. However, Yura et al (1980) made the
first major attempt to collect and critically examine the test results obtained
world-wide. This resulted in the development of a set of formulae for axial and
moment loaded T, Y, DT and K joints which largely form the basis of current
recommendations of the API code (1989). These formulae are derived to form a
lower bound to a carefully selected data-base of 137 test results. Similarly in 1982
Billington et al published a paper giving a series of formulae for the ultimate strength
of axial and moment loaded T, Y, DT and K joints which later appeared as guidelines
of the Underwater Engineering Group (1985) for the design of tubular joints. The
data-base consisted of 207 test results which included most of the data used by Yura
et al (1980). The latest UK DEn. (1990) guidelines on ultimate strength design of
joints are developed on similar lines to U.E.G. recommendations but use a more recent
data-base of test results. The limit-load formulations given in the following are based
on the background document, Dept. of Energy (1989b, 0TH 308).
4-4

BRA

CHO

DI
L
4

D D 2T D T

Fig.4.1: Nomenclature for tubular joint geometrical parameters

Strength prediction models are first presented for each loading mode separately and
then the effect of combined loading is considered. For a given loading mode, the
strength of a joint depends on the 'joint type', a classification which considers both the
geometrical configuration of planar brace and chord members meeting at a node and
load-balancing among brace members, see D.En. guidance notes (1990). The
geometrical notations used and non-dimensional parameters are defined in Fig.4. 1.

4.2.1 Compression Loaded Joints:

Typical load-deformation curve for joints loaded by axial compression is shown in


Fig.4.2. During testin.g it was observed that first yield generally occurs at ihe
'hot spot' locations. Under axial compression, failure of a joint occurs by buckling
of the chord wall after significant plastic deformation. The mean strength formulae
for compression loaded joints can be given as, see D.En. (1989b, 0Th 308)
4-5

AXIAL LOAD
400

7-Joint
300 Elastic Tension
Behaviour
of Brace
P
(kN)
DT-Joint
200 Tension

______ T-Joint
- Compression
100
DT-Joint
Compression

0'
0 5 10 15 20 25 30
AXIAL DEFORMATION (mm)

Fig.4.2: Typical axial load-deflection curves for tubular joints, Yura et al (1980)

For TA' and K/YF joints:

FT2
P - (2.37 +23.60) ka Qg
sin8

- 1.67-O.86f For C^O.6

- 1.0 For >O.6


(4.la)

0.3
Q13 - For f3>0.6
3(1-0.8333)

- 1.0 For f3^0.6


4-6
For DT and X joints:

FT2
P - (2.98+15.4513)Qk0 (4.lb)
sine
where F is the yield strength of the chord plate.

The formula for strength of T and Y joints is based on a data-base of 60 T, Y and


large gap K/YT joint test results available until 1986, with beta ratios varying from
0.18 to 1.0. The joint strength is strongly dependent on the parameter 1 3 while y is
seen to have no influence. The joint strength increases considerably for (3 ratios
beyond 0.6 and the factor Q is introduced to account for this fact. Some of the more
recent tests conducted by Wimpey Offshore Engineers, see D.En. (1989a, 0TH 297),
show an influence of y on joint strength at high 1 3 values. However, part of the
increase in joint strength observed in these tests could be due to the chord end
conditions used. The parameter a may also have some influence on joint strength as
for higher values of cx (i.e. longer chord span length), the extreme fibre chord stresses
increase due to chord beam behaviour, resulting in a reduced bending strength and
hence a reduced joint capacity. However, the existing data-base is insufficient to
quantify the effect of? and a parameters. Note that API RP 2A also does not
account for y and a effects.

For DT and X joints the formula indicates a reduction in joint strength, by about 25%,
compared to corresponding T and Y joint configurations. The term ka accounts for
the increase in intersection length for Y and X joints and the consequent increase in
joint strength. The API RP 2A formulae, however, do not use the ka term, possibly
because of lack of data at the time these guidelines were drafted.

For a joint to be designed as a K joint, the perpendicular component of the


compressive force in one brace should be balanced by the component of the tensile
force on the second brace. Thus, YT joints satisfying this requirement can be
designed as K joints. The strength of K joints is seen to be strongly dependent on the
gap between the two braces. As the gap reduces, the strength of one brace increases
4-7
due to the stiffening effect of the other brace. The failure of the joint is governed by
the stress state in the chord plate in the gap. Failure usually occurs at the compressive
brace and hence axially loaded K joints are classified under compressive joints. The
tension loaded brace is separately checked as a Y joint.

4.2.2 Tension Loaded Joints:

Typical load-deformation curve for tension loaded joints is given in Fig.4.2. Under
high tensile loading of the brace, the chord wall around the brace undergoes plastic
deformation showing necking and elongation of the chord. As the load increases, a
crack appears due to tearing of the material near the hot-spot region but the joint
continues to carry higher loads until the joint finally fails by the separation of the
brace from the chord. The mean strength formulae for joints under axial tension can
be given as, D.En. (1989b, 0TH 308)

For T and Y joints:

FT2 (4.2a)
P - (11.70+32.2613) ka
sinO
For DT and X joints:

FT2
Pk - (9.20 + 22.63 J3) (4.2b)
'a
sine

The D.En. formulae for tension loaded joints are based on the results of 18 T/Y and
16 DT joint tests and predict the maximum load which a joint can take. It can be
seen that strength of a joint under axial tension is significantly higher than under axial
compression (by a factor of about 2!). It is also important to note the considerable
scatter in test results. The API RP 2A formulae do not indicate appreciable increase
in capacity under tension loading compared to compression loading as the design load
is defined corresponding to the onset of tension crack, see Yura et al (1980).

4-8

ODIDING ODIT (kN-m)

200

160
In-Plans B.ndIn4

100

fOut-of-Plane Dendin4

60/

ØL

0 0.06 0.1 0.16 0.2 0.26 0.3
ANGULAR D0RKA110N (Radians)

Fig.4.3: Typical moment-curvature curves for tubular joints, Yura et al (1980)

The strength of DT and X joints is reduced by a factor of about 0.75 compared to the
corresponding T and Y joints. The D.En. formulae use the factor ka to account for
the increased length of the intersection in Y and X joints while the API formulae do
not use these factors.

4.2.3 In-plane Moment Loaded Joints:

Typical moment-curvature relationships for joints subjected to in-plane bending


moment axe given in Fig.4.3. Failure of a joint under in-plane bending typically
occurs due to excessive plastic deformation of the chord wall on the compression side
followed by tearing of the material on the tension side of the chord. The mean
strength formula for all types of in-plane moment loaded joints is given by D.En.
(1989, 0TH 308), as
4-9

For all joint types:



M - FT 2 d (6.2013 -0.27)/ (4.3)

The value of the moment at the chord-brace surface intersection was used in reporting
the test results and in deriving the above equation. The use of moments at the
intersection of chord-brace axes, as usually obtained from a global structural analysis,
will therefore be conservative. Note that the full value of the applied moments is used
instead of the perpendicular component as it is felt that under in-plane bending the
applied moment is resisted in total by the joint.

The D.En. formula is based on the results of 31 T joint tests and shows an influence
of both (3 and 'y parameters. The formula does not account for the influence of the
gap parameter and intersection length parameter due to paucity of data. Recent tests
conducted by Wimpey Offshore, see D.En. (1989a, 0TH 297), confirim that the
influence of the gap parameter on strength of K joints is negligible. However, an
influence of moment direction was observed in K joints showing a slight increase in
joint strength when the moments in the two braces were applied in an opposite sense
(i.e. causing stresses of the same sign in the chord plate in the gap region). A definite
influence of the relative section factor k was observed for Y joints at low (3 ratios.
The current D.En. and API guidelines do not allow the use of this parameter.
However, for K joints the use of the k term is not considered appropriate, especially
at high (3 values, as some of the benefit gained from the increased section is offset by
the detrimental effect of stress field interactions in K joints.

4.2.4 Out-of-plane Moment Loaded Joints:

Typical moment-curvature relationship for out-of-plane moment loaded joints is shown


in Fig.4.3. For this mode of loading it was observed during testing that chord wall
distorts in the vicinity of saddle points resulting in lower joint stiffness and strength.
Under continued loading the deformation increases rapidly without significant increase
4-10

in loading until fmal failure occurs by tearing of the chord material on the tension
side. In many cases tests had to be stopped at very high deflections due to limitations
of the test rig. The mean strength of joints subjected to out-of-plane bending can be
given as, see D.En. (1989b, 0Th 308),

ForT, Y and Kjoints:

FT2d (
M - _____ 1.88 +8.643) Q (4.4a)
sinO

For DT and X joints:

FT2d
M - ( 1.88+8.64)v/
(4.4b)
sinG

The formula for T/Y joints is based on test results of 18 joints and show a significant
increase in joint strength for f>0.6 which is accounted through the term.
Additional tests conducted at Wimpey Offshore on K joints do not show any influence
of the gap parameter and relative section factor kb,, D.En. (1989a, 0TH 297). Only
two test results are at present available for DT and X joints subjected to out-of-plane
bending. The reduction in capacity for double joints compared to single joints is to
be expected, particularly at high f values, due to increased interaction of stress fields
at the saddle locations. This is accounted through the IQ term in the equation.

4.2.5 Influence of Chord Loads on Joint Strength:

The presence of forces in the chord inducing stresses at the intersection with the brace
will have a significant influence on the ultimate strength of the joint. When a
compressive load in the chord is associated with a tension load in the brace the joint
strength should supposedly increase. However, when compressive, in-plane moment
or out-of-plane moment loads in the brace are associated with axial compression or
moment loads in the chord causing compressive stresses in the vicinity of the
intersection, the strength of the joint reduces considerably.
4-11

For joints subjected to axial compression through the brace, the reduction in strength
can be attributed to the fact that chord compressive stresses reduce the local bending
strength of the chord plate at the joint. In addition, once the chord wall deforms a
P-8 secondary moment develops local to the joint. Test results, however, show that
reduction in capacity is not appreciable for joints with high values (>0.8). This
could be because in such joints the applied brace loads are resisted by the chord
mainly by a membrane action and little deformation occurs until the ultimate load is
reached. In the case of chord tensile stresses the P-6 term acts as a restoring moment
thereby preventing significant loss in capacity as evidenced in some tests.

The effect of chord load on in-plane moment loaded joints is seen to be more severe
than the out-of-plane moment loaded joints since the chord resists in-plane brace loads
near the crown locations where significant deformation takes place. In the case of
out-of-plane moment loaded joints chord resistance is mobilised at the saddle locations
in a predominantly membrane mode.

The effect of chord loads is accounted through a chord load reduction factor Q 1 which
gives the ratio of ultimate strength of a joint with chord loads to joint capacity without
chord loads. This is related to the y ratio of the joint, which reflects the relative
flexibility of the chord member, and a 'chord utilisation factor' U, which expresses the
severity of chord loading. The D.En. (1989b, 0TH 308) gives the following formulae
for the chord load factor.

Axial Loading:

Q 1-0.030'yU2 (4.5 a)

In-plane Bending:

QI - 1-0.045yU2 (4.5b)
4-12

Out-of-plane Bending:

Q1 - 1-O.O21yU2 (4.5c)

U - /(O.23DP)2+M2+M:
O.72D 2 T F

P - chord axial force


M - chord in-plane moment
M0 - chord out-of-plane moment

These reduction factors should be multiplied with the ultimate strength computed from
the equations given previously for individual modes of loading.

4.2.6 Strength of Cracked Joints:

Due to cyclic loading of waves, tubular joints in offshore structures can undergo
deterioration due to fatigue. The presence of a fatigue crack of significant size will
reduce the static strength of a tubular joint. Unfortunately, no published data is
available on the strength of joints with fatigue cracks. An approximate method for
determining the strength of damaged joints is attempted here.

The term FT2 in the strength prediction equations for intact joints presented above
is based on a theoretical ring analogy and can be related to the plastic moment of a
ring model of unit width. The strong dependence of joint capacity on parameter 1 and
ka indicate that joint resistance is mainly derived from the chord material along the
intersection with the brace member. The presence of a crack in the chord plate
reduces the volume of the material resisting the applied forces and hence the capacity
An equivalent plate model of a cracked joint can be considered as shown in Fig.4.4.
eM1
4-13
4P
d 4 1 M0

Ttd
,,weld
_411r

—'1 'eqv

Fig.4.4: Equivalent plate idealization of a cracked tubular joint

A preliminary study, using finite element analysis, was conducted by Framatome of


France, in association with the author of this thesis. In this study an attempt was
made to correlate the elastic stress distribution of an intact joint to its static strength.
It was observed that membrane action of the chord plate plays a major role in
balancing applied brace forces, see Heliot and Benoit (1989). If the membrane action
of the chord plate increases an increase in joint strength could be expected. This
observation correlates well with experimental results showing higher capacity for
joints with high ratios which also have a high membrane stress component, see
discussion in section 4.3 later in this chapter. The presence of a crack of significant
size also reduces the local bending stiffness at the crack location thereby relieving
some of the bending stress. Based on these observations, only the membrane
component is assumed to be effective in determining capacity of a cracked joint.
4-14

From the equivalent plate idealisation of a cracked joint, Fig.4.4, subjected to pure
membrane stress, a ratio of the strength of the cracked joint to its intact strength can
be obtained as

- strength of cracked joint


strength of intact joint
- (leqyT_O.5irac)/lq.,,T
(4.6)
a - crack depth
c - crack semi-length
- equivalent length

The equivalent length L of the equivalent plate represents the length of the
brace/chord intersection over which the stress distribution can be assumed to be
uniform. This will depend on the stress distribution along the intersection and is
expected to vary depending on such factors as loading mode, joint type etc. For axial
loading nearly full length of the intersection may be used while for in-plane and
out-of-plane bending loads only a part of the intersection near the crown or the saddle
points should be used. In the absence of concrete information the equivalent length
is assumed as the full intersection length.

4.2.7 Joints Under Combined Loading:

When a joint is subjected to the combined action of axial loading and bending
moments through the brace, the interaction of stress fields is likely to have a
detrimental effect on joint strength. Based on the results of 74 joint tests D.En.
(1989b, 0TH 308) defines the failure surface for combined loading by an interaction
equation of the form

( Pa (Ma ' fM
I I + I ' + I - I - 1.0 (4.7)
ITi iVI
N) IM I
¼
N,JPB
4-15

where P and M represent applied axial load and bending moment respectively. It is
interesting to note that the combination of axial loading with in-plane moment is not
that critical for joint capacity. This is to be expected as significant proportion of the
joint resistance under axial loading is mobilised at the saddle locations while in the
case of in-plane moment the resistance is developed around the crown locations,
thereby causing little interaction of stress fields.

The capacity models discussed so far relate to only simple joints in which brace and
chord members are all in a single plane. At present very limited test data is available
for complex, multi-planar and overlapping joints. References U.E.G (1985) and D.En.
(1989b, 0TH 308) offer general guidelines for the treatment of such joints. Since all
the strength models presented in this section are developed from experimental results
which exhibit considerable scatter, uncertainties will be involved in the strength
estimates made using these models. Quantification of model uncertainties associated
with the above limit-load models is presented in chapter 5.

4.3 STRESS DISTRIBUTIONS IN TUBULAR JOINTS

Because of the tubular member cross-sections the intersection detail is of a


complicated geometry even for simple joints. This complicated geometry, along with
the presence of the weld at the intersection, causes major distortions of the nominal
stress field in each of the brace and chord members framing into the joint. For fatigue
and fracture assessment of tubular joints a knowledge of the detailed stress distribution
at the joint is required.

The stress distribution at a tubular joint is governed by four main causes as follows:

1. Stresses due to basic structural response of the brace and chord members of
a joint to applied loading, termed 'nominal stresses'. These stresses can be
obtained from the nominal forces obtained from a global structural analysis.
4-16

2. Stresses arising due to the deformation of the brace and chord walls at the
intersection. The difference in axial stiffness of the brace and bending
stiffness of the chord wall, which varies along the intersection, causes major
distortions of the intersection resulting in maldistribution of the stress field.
These are referred to as 'geometric stresses' since the magnitude and
distribution of these stresses is influenced by the overall geometry of the joint.

3. 'Notch stresses' are the result of geometric discontinuity of the tubular walls
at the weld toe where an abrupt change in section occurs. These stresses are
influenced by the geometry of the weld and are relatively independent of
overall joint geometry.

4. 'Residual stresses' are those introduced during fabrication and welding of the
joint. They remain constant around the intersection and are independent of
joint type.

Determination of nominal stresses using a deterministic or spectral response analysis


has already been discussed in chapter 2. In the following, methods to determine
geometric, notch and residual stresses are discussed.

4.3.1 Geometric Stresses:

Geometric stresses can be determined from a stress analysis of a joint in which the
detailed geometry of the chord and brace members at the intersection is modelled. A
number of methods are available for stress analysis of tubular joints. Finite element
methods based on thin shell theory are widely used for determining geometric stress
distribution. The accuracy of results depends on the mesh size and type of element
used and convergence of results shou'd be carefully studied. Mesh generation usually
is very time consuming and specialised mesh generators for tubular joints are essential.
Where possible symmetry conditions should be used to reduce the cost of
computation. Experimental stress analysis methods are used where high accuracy is
4-17
required or, as in most cases, to establish the accuracy of numerical methods. Steel
models are used when the effects of weld geometry is to be properly reflected.
Acrylic or araldite models are used for parametric studies as they are cost effective.
Strain measurements are made using electrical resistance strain gauges and brittle
lacquers or using the technique of photo-elasticity. Each of these techniques are quite
specialised and require considerable experience.

Considering the total number of joints in a structure and the cost involved in stress
analysis of one joint using either numerical or experimental methods, these methods
are clearly not practicable for routine use. Empirical formulae have therefore been
developed from parametric studies covering different joint types and geometries.
These studies have used either fmite element methods or experimental methods based
on acrylic models. In view of their simplicity and cost-effectiveness , the use of
parametric equations is considered appropriate in a system reliability analysis where
a large number of joints are involved. Parametric solutions for stress analysis of
tubular joints is discussed in the following sections.

4.3.1.1 Stress Concentration Factors:

In determining geometric stresses at a joint it is customary to use a term called 'Stress


Concentration Factor' (SCF) which expresses the ratio of the stress at the weld toe of
a joint to the nominal stress in the brace member. Parametric equations normally
determine SCFs corresponding to saddle arid crown locations at a joint under axial
loading, in-plane bending and out-of-plane bending in the brace member. The value
of SCF will be generally different for the brace and chord side locations of the joint.

Over the years a number of SCF formulae have been reported in the literature of
which the early comprehensive works include Kuang ci al (1975), Wordsworth and
Smedly (1978) and Wordsworth (1981). Later U.E.G (1985) proposed some
modifications to Wordsworth and Smedly formulae which extended the range of
application and removed some of the un-conservatism in these formulae. More recent
formulae for SCF are from Efthymiou (1988) and Hellier, Conolly and Dover (1990).
4-18
All these references give different formulae for SCFs which unfortunately differ
considerably. The discrepancy between formulae is largely due to the method of
analysis used, choice of location at which SCFs are computed, chord end fixity
conditions and range of joint geometries studied in developing the equations.

The Kuang formulae were developed from finite element analysis using "fixed"
boundary conditions for chord ends. The joint geometries considered ranged from
O.33^f3^0.75 but all joints had an a value of 7. The computed SCFs relate to the peak
stress at the intersection of chord/brace mid-surfaces. The Wordsworth and Smedly
formulae, often referred to as Wordsworth formulae, on the other hand were developed
from Acrylic model tests with "pinned" chord end boundary conditions. The U.E.G
formulae and Efthymiou formulae correspond to the peak stress at the weld toe
obtained by linear extrapolation of the principal geometric stress in the vicinity of the
joint. This is also the agreed definition for determining 'hot-spot' stress used in
calibrating S-N curves for fatigue analysis.

All the above formulae exhibit considerable scatter in their predictions when compared
with experimental results. Of the considered formulae the formulations by Hellier
et al give the best correlation with experimental results. However, these formulae are
limited to T and Y joints only with 1 3 values less than 0.8. Only the U.E.Ci. formulae
and Efthymiou formulae cover T/Y, YT/K and DT/X joints and are applicable over
a wider range of joint geometrical parameters. The Efthymiou formulae also cover
KT joints and K joints with overlap and consider different types of loading conditions
such as balanced loading in two braces, loading in one brace only and unbalanced
loading in both braces etc. In addition the formulae for T/Y joints explicitly account
for chord end fixity and short chord effects for a<12. The format of equations allow
for easy manipulations to obtain hot-spot stresses in multi-planar joints under
combined loading which is close to the practical situation. Thus the Efthymiou
equations represent the most comprehensive and rigorous formulation for SCF
available at present. The various equations for different types of joints and loading
conditions are tabulated in Efthymiou (1988) and are not repeated here. The model
uncertainty associated with these equations is discussed in chapter 5.
4-19
Although the SCF formulations discussed above are simple to use they suffer from
some drawbacks. First the presently available formulations cover only simple planar
joints. The formulations for K, KT and X joints generally assume balanced loads in
the brace members at a node. However, in practice many types of complex and
overlap joints are used and braces at a node can be in more than one plane.
Moreover, the ratio of axial loads and moments in different braces at a node vaiy for
different wave approach angles, wave period and for the same wave they vary as the
wave passes through the structure. This means joint type identification and SCF
calculations have to be repeated for every wave position. This will increase the
computational cost considerably and this approach is usually not followed in practice
and the computed hot-spot stresses may be in error. Recently Efthymiou (1988) has
proposed an 'influence function' approach which avoids repetitive calculation of SCFs.
The approach is essentially that of a linear superposition of hot-spot stress
contributions from each of the braces meeting at a node. Influence function
formulations for hot-spot stress in different types of joints under different loading
conditions have been derived in Efthymiou (1988). The proposed formulations can
be easily extended to multi-planar joints. A comparison of this approach with
conventional SCF approach has shown that multi-planar effects are usually small and
conventional SCF approach generally yields conservative estimates of hot-spot stresses
except in some cases. The influence function formulation could become the standard
practice in the coming years.

4.3.1.2 Stress Distribution Around the Intersection:

The SCF formulae presented above predict peak stresses at only the saddle and crown
positions around the intersection separately for each mode of brace loading. It is
observed that under in-plane bending of the brace highest stresses occur at the crown
locations with zero stresses at the saddle positions. While under out-of-plane bending
the highest stresses occur near to the saddle locations with zero stresses at the crown
locations. Under axial loading highest stresses occur around the saddle positions with
non-zero stresses at the crown locations. Thus for individual modes of loading it is
easy to determine the 'hot-spot' stress, defined as the highest geometric stress
4-20

anywhere around a brace/chord intersection. However, under combined loading it


becomes necessary to calculate combined geometric stresses at a number of points
around the intersection. Normally, combined stresses are computed at 8 or 12 points
and the highest value of these is taken as the hot-spot stress. It is thus necessary to
determine stress or SCF variation around the intersection for each mode of loading.

Unfortunately, published information on SCF variation around the intersection is


lacking and simplified assumptions are often made in design practice. Gulati et al
(1982) propose a linear variation of SCF values between the saddle and crown
positions for the axial load case and a simple sinusoidal variation for the moment load
cases. Based on comparison with experimental results on steel models Dharmavasan
and Dover (1984) have suggested the following interpolation function:

Kg (v) - C() [ K8 cos2 W +KgsSfl2N1/C(90')1


'2 2
C (W) - 2sinO/d
(4.8)
Kgc - stress concentration factor at crown
Kg3 - stress concentration factor at saddle

w - angular position from the crown

The combined geometric stress at each location around the joint is determined by
superposing stress contributions from each mode of loading.

4.3.1.3 Through-Thickness Stress Distribution:

A knowledge of the hot-spot stress is usually sufficient for a S-N type of fatigue
analysis but for a fracture mechanics analysis information on the stress distribution
through the thickness of the plate containing a crack (usually the chord) is required.
The hot spot stress distribution through the thickness of the chord wall is often
assumed to be linear, comprising a membrane component and a bending component.
The through-thickness distribution is described in terms of a 'Degree-of-Bending"
(DoB) factor giving the ratio of the bending stress to the total stress. This ratio is
4-21
found to depend on the mode of brace loading and joint geometrical parameters and
will vary around the intersection.

The parametric formulae given by Conolly et a! (1990) are the only published
information available at present for determining DoB factors. These formulae were
developed from extensive finite element analyses but cover only T and Y joints. The
formulae predict the DoB factors at the saddle and crown locations of a joint under
brace axial loading, in-plane bending or out-of-plane bending. Unfortunately, no
formulae are available to determine the variation of this factor around the intersection
and for this reason it is assumed to follow the variation of total hot-spot stress at the
outer plate surface. More studies are needed to develop formulae to compute this ratio
for other types of joints as this factor has a significant bearing on the accuracy of
fatigue lives computed from a fracture mechanics analysis.

4.3.2 Notch Stresses:

In defining the geometric or shell hot-spot SCF the effect of weld geometry at the
joint is not considered. The presence of the brace member and the weld bead increase
the stiffness of the chord wall locally causing further stress concentrations. This
additional stress concentration is accounted through a 'notch stress concentration
factor' K, defined as the ratio of the maximum total stress at the weld toe to the
geometric stress at the location.

The notch stress concentration factor can be determined by finite element analysis or
using photo-elastic methods. Elliott and Fessler (1986) have carried out photo-elastic
stress analysis of tubular joints to determine K. They observed that K is affected
by the weld size, weld angle and weld toe radius but relatively uninfluenced by the
overall geometry of the tubular joint. The notch stress concentration factor increases
proportionately with the weld angle but decreases rapidly as the weld toe radius
increases. Burdekin et al (1988) have reported empirical formulae for K1 based on
2-D finite element analysis of cruciform butt welds which can be stated as
4-22

K-
($)O.353
K, - 1.206(4))° (1 IT
)O.11I
for 1IT<l ; p-0
(4.9)
K 4 - 1.694 (4) )O.271
(1IT ) O.037 ($)° for llT>1 ; p-0

- 1 +0.23(p/T)°6 for 4)-45°

where L is the weld leg length, 4) is the weld toe angle and p is the weld toe radius.

The notch stress distribution in a tubular joint will be 3-dimensional in nature and the
use of above equations for tubular joints is expected to yield conservative results.
From experimental measurements it is observed that weld notch stresses fall off
rapidly away from the weld toe on the chord surface and also in the thickness
direction below the weld toe. The extent of notch zone is found to be about 0.35T
in most cases but in some cases can extend up to 0.46T for weld angles greater than
60. Niu and Glinka (1987) have given empirical formulae for through-thickness
variation of notch stresses for T-butt weidments subjected to nominal pure bending
loading through the plate which can be stated as

/2+i'
KnSgI(x 1h/2 1 312 ji
S(x) -
2)
+ !(
2p +zc[) ][i-]
2LP
C0 - 1 .4594) 2 _3 .0564) +2.4 for (it/6)^4)^(it/3) (4.10)

C1 - 1

xo -
p

where S5' is the geometric stress, K is the notch stress concentration factor at the
surface, x is the distance from the plate surface and X is the distance to the neutral
axis. Thus the contribution of bending is controlled by the parameter;. It is seen
that weld toe angle has a greater influence on the through-thickness distribution of
notch stress while the weld toe radius has a considerable influence on K, Niu and
Glinka (1987). A similar expression for membrane loading is not available.
4-23

4.3.3 Residual Stresses:

In addition to the stresses caused by the external action of loads, residual (or
locked-in) stresses are present in tubular joints. These stresses arise from fabrication
and welding procedures and are generally self-balancing over the brace-chord
assembly at the joint. The large plastic deformations induced during welding and
unequal rate of cooling of different parts result in thermal stresses which remain
locked in the body of the plate. These stresses can be removed by re-heating the weld
and the surrounding parent plate and cooling under controlled conditions, a process
known as stress relieving. In practice, however, not all joints in a structure are stress
relieved and even when carried out is not fully effective and thus residual stresses of
significant magnitude are always present.

The magnitude and distribution of residual stresses in a welded component depends


on such factors as weld type, overall geometry of the component, welding heat input,
velocity of weld passes, yield strength of the material etc. and their theoretical
prediction is very complicated. A number of experimental techniques, such as
bore-hole drilling, slicing and X-ray diffraction etc. axe used for measurement of
residual stresses. These methods are expensive and time-consuming and are not
practicable for use on all joints in a structure. However, residual stresses have a
significant influence on fatigue and fracture strength of tubular joints and some
simplified way of accounting for them is necessary. Design codes often recommend
a conservative value of tensile yield strength to be assumed for residual stresses
uniformly distributed through the plate thickness close to the weld.

Experimental measurement of residual stress distribution has been reported by Payne


et al (1986) and Porter-Goff (1988). Through-thickness distributions of transverse
stress (normal to the weld) and longitudinal stress (parallel to the weld) representative
of conditions at the weld toe for tubular T joints is shown in Fig.4.5.
4-24

Stress (MP)

I Stress (HPa)

TRANSVERSE WNGITUDINAL

Fig.4.5: Typical through-thickness residual stress distribution, Porter-Goff (1988)

From these limited studies following inference can be made about the distribution of
residual stresses in tubular joints:

1. Since the weld is present only on the outer surface of the chord plate, the
residual stress close to the weld will be tensile in nature and may well reach
yield stress in magnitude.
2. Residual stresses fall off rapidly in thickness direction to less than half the
value within 10 mm from the surface and may well become compressive at
mid-thickness of the plate. They may again reach small tensile values at the
inner surface of the plate. Thus the through thickness distribution consists of
a predominant bending component and a small membrane component.
3. Considerable scatter was observed in magnitude of surface residual stress
values during experimental measurements and coefficient of variation of up to
25% can be expected.
4. The assumption of uniform residual stress distribution through the plate
thickness is not realistic and can lead to highly conservative results.
4-25
The through-thickness residual stress distribution in this thesis is therefore idealised
as yield value in tension at the plate surface varying linearly through the thickness of
the plate. From the limited data available it is not possible to quantify the extent of
bending and membrane components. In the absence of precise information it is
reasonable to assume that the degree of bending factor for residual stresses will be the
same as for applied geometric stresses.

The idealized combined stress system, comprising geometric stress, notch stress and
residual stress, through the thickness of the chord plate of a tubular joint adjacent to
the weld toe is shown schematically in Fig.4.6.

II St0t

(a)

II
Sm Sb(a) "I-i

I"
1-a/T

(b)

Fig.4.6: Through-thickness stress distribution at a tubular joint


(a) Geometric stress (b) Residual stress
4-26
4.4 FATIGUE FAILURE OF TUBULAR JOINTS

Fatigue in offshore structures occurs due to the oscillatory nature of ocean waves and
is normally confined to tubular joints in the structure which are regions of high stress
concentrations. Fatigue is a particularly serious problem for North Sea structures due
to frequent storm conditions. Fatigue in tubular joints is an extraordinarily
complicated process involving many factors. Fatigue failure is characterized by
initiation and propagation of cracks and can be best understood by studying the
mechanics of crack growth.

This section discusses the fatigue characteristics of offshore structures and the S-N and
fracture mechanics approaches to fatigue life estimation. The available solutions for
stress intensity factors are reviewed and difficulties involved in the fatigue life
estimation of tubular joints highlighted. A simple fatigue crack propagation model for
tubular joints is developed and its performance is compared with experimental results.

4.4.1 Fatigue Characteristics of Offshore Structures

Fatigue in offshore structures is a complex process and a fully satisfactory solution for
estimation of fatigue thmage has not yet been developed. However, some important
characteristics of fatigue behaviour of offshore structures have been established and
it is essential that these be realistically represented in an analysis.

4.4.1.1 Factors Affecting Fatigue Behaviour:

Fatigue behaviour is influenced by many factors which can be grouped as:

(a) Loading: Fatigue is a cumulative, time-dependent phenomenon. Each cycle


of load will cause some fatigue and thus the magnitudes of all load cycles over
the service life of the structure characterize the fatigue loading. In offshore
4-27

structures the predominant fatigue loading is due to ocean waves which means
a detailed description of the long-term wave induced stress cycles is required.

(b) Joint Geometry: Since fatigue crack advance depends on the cyclic stress
range at the very tip of the crack, the stress analysis of joints should account
for the stress raising effects of the overall geometry of the joint and the local
geometry of the weld at the crack site.

(c) Weld Defects: Fatigue cracks in tubular joints originate from weld defects
such as inclusions, undercuts, lack of fusion etc. which are to some extent
unavoidable. Though there may be many defects in the joint only those
defects which are present in regions of high stress concentration develop into
fatigue cracks. Thus weld defect size and their distribution in the body of the
weld have an important influence on fatigue life and unfortunately information
on this is usually not available for all joints in a structure.

(d) Material Property: Crack propagation depends on the cyclic elasto-plastic


response of the material near the crack tip. Material properties describing
crack growth behaviour are obtained from experiments combined with an
assumed crack propagation model and generally a large scatter in results exists.

(e) Environment: A sea water environment causing corrosion of the material is


seen to accelerate crack growth. The mechanism of corrosion-fatigue is not
well understood and is observed to depend on many factors such as chemical
content of sea water, temperature, current, frequency of loading, effectiveness
of cathodic protection etc. which are difficult to determine precisely.

(f) Others: The presence of residual stresses introduced during welding, mean
stress due to static loads and the sequence of load cycles in a variable
amplitude loading are observed to have some influence on fatigue life.
4-28
4.4.1.2 Observed Crack Propagation Behaviour in Tubular Joints:

Fatigue crack growth has been recorded by careful measurements during fatigue tests
on tubular joints. Observations of crack growth have been reported for T joints by
Martin (1978), for Y joints by Dover and Dharmavasan (1982) and for
non-overlapping K joints by Wylde and McDonald (1981). The general features of
crack growth are similar in most joints.

Fatigue cracks generally initiate at several positions around the welded intersection
and are clearly visible within about 10% of the total fatigue life. Crack initiation
points are closely grouped for joints subjected to bending moments through the brace
while they are widely distributed under axial loading as a large extent of the joint
intersection is subjected to high stresses. Multiple cracks soon join up to form a
single dominant fatigue crack. Initiation usually occurs in the chord plate for most
joints due to the higher notch stresses at the weld toe.

The use of crack microgauges based on A.C. potential drop technique, beachmarking
or observation of fatigue-fracture surface have helped to study the crack growth
behaviour in the depth (thickness) direction. It is found that after initiation, fatigue
crack begins to grow both in the depth direction and in the circumferential direction
along the weld toe. After an initial delay, the crack depth increases approximately
linearly with the number of applied stress cycles until the crack breaks the chord wall
at its deepest point. The crack propagation rate thus remains fairly constant and does
not accelerate as the crack becomes deeper which is quite unlike the behaviour found
in flat plates, Dover et al (1988).

The crack growth in the circumferential direction at the surface is rather more
complicated and exhibits considerable scatter in test results. This could be because
of the changing weld geometry and stress gradient which depends on joint type, joint
geometry and mode of loading. Crack propagation has been found to be faster in
axially loaded joints compared to out-of plane bending. The crack aspect ratio,
defined as the ratio of crack depth to semi-crack length, also changes as the crack
4-29
grows, increasing from a value of about 0.02 at initial stages of growth to values of
between 0.1-0.2 at later stages. The aspect ratios for different joints show large
scatter, see U.E.G. (1985, Vol 2). When the crack penetrated the chord wall at its
deepest point, the surface crack was observed to extend to about a third to half the
intersection circumference. In the later stages of crack growth, the surface crack was
found to branch out into the chord wall away from the weld toe.

Once the crack penetrates the chord wall at its deepest point, the ligament over the
remainder of the crack length was found to break through in a relatively small
number of stress cycles to form a rectangular through-thickness crack over almost the
surface crack length. After this point the surface crack growth Tate was found to
increase rapidly in most cases, Wylde and McDonald (1981).

In a few cases joint stiffness measurements were made in terms of the actuator tip
displacement. It was observed that the displacement remained virtually constant until
through-thickness cracking occurred. After this stiffness reduced gradually and it was
only in the last 10% of the life that the stiffness deteriorated significantly. In some
cases the displacement fmally became so high that tests were limited by the actuator
stroke and in such cases this point was used to define "end of test".

4.4.2 Fatigue Analysis Using S-N Approach

This is the most commonly used method of fatigue analysis in the offshore industry.
In this method the fatigue strength of a component is characterized in terms of an S-N
curve which is a plot of stress range (as ordinate) versus number of stress cycles to
failure (as abscissa), both plotted on a logarithmic scale. All the offshore design
codes specify S-N curves for the fatigue design of tubular joints which in early years
were derived from fatigue tests on plated joints and a limited number of small scale
tubular joints. However, during the UKOSRP (see D.En. 0TH 88 262) and ECSC
(1987) research programmes a large number of tubular joints of various sizes were
tested to study the influence of several factors. This resulted in the new fatigue design
4-30

curves recommended by the Dept. of Energy (1990), see D.En. (1984) which forms
the background document for these guidelines. The UKOSRP-II project which
concluded in 1987 (see D.En. 0TH 87 225) produced new data on fatigue strength of
tubular joints which led to the revision of D.En. guidelines which are currently under
review by the industry, see Reynolds and Sharp (1990).

The approach followed in D.En. (1984) was to first establish a basic S-N curve based
on joints tested in air with section thickness of 32mm. Modifications are then
suggested to account for the influence of other factors on fatigue life. The following
discussion on S-N curves mainly follows D.En. guidelines including recent revisions.

4.4.2.1 Basic S-N Curve:

The basic S-N curve is established for tubular joints of a single thickness for air
conditions. The 1984 D.En. guideline used 32 mm joints as the reference thickness
while in the new revisions 16 mm joints are used as the reference thickness as
considerably more data are available for this thickness. The new curve is derived
from results of a total of 59 tubular joint fatigue tests conducted in air which cover
a variety of joint types and loading modes. The S-N curve has two parts with
different slopes which can be expressed in the form

log(N) - log(K1 )-mJ log(S) for N^107


(4.11)
log (N) - log (K2) - m2 log (SB) for N> i07

The inverse slope m 1 for the first part is taken as 3 while for the low stress-range part
m2 is taken as 5. The mean values of log(K1 ) and log(K2) are obtained from
regression of test data. The design S-N curve in both the revisions corresponds to
mean minus two standard deviations of log(N) of the 16mm data-set. However, for
the purposes of reliability analysis presented in this thesis only the mean value
S-N curve is relevant.
4-31
The basic S-N curve as derived above is calibrated with measured hot-spot stress
which is defmed as the greatest value of the geometric stress anywhere around the
brace/chord intersection. The geometric stress, as defined before, is the extrapolation
to the weld toe of the linear part of the principal stress variation near the joint. All
tests used in deriving this S-N curve were as-welded joints without profile control and
post-weld heat treatment and were tested under constant amplitude loading. Thus the
effects of weld geometry, residual stresses and through-thickness stress variation is
implicitly included in the S-N curve. It is recommended that variable amplitude
loading can be treated using Miner's linear damage summation. The effect of other
factors is accounted through appropriate corrections to the basic S-N curve as below.

4.4.2.2 Effect of Plate Thickness:

A number of fatigue tests have shown that, for a given cyclic stress-range, fatigue life
decreases as the plate thickness of a joint increases, a phenomenon commonly called
"thickness effect". The Dept. of Energy's latest revisions suggest a reduction in
fatigue strength to be calculated using, see Reynolds and Sharp (1990)

t (4.12)
S - SB
t

where
S is the fatigue strength of the joint under consideration
SB is the fatigue strength obtained from the basic S-N curve
t is the actual thickness of the plate under consideration
t B is the reference thickness for the basic S-N curve (16mm)
is the thickness correction exponent (0.30)
The thickness correction exponent in the 1984 guidelines was 0.25 with a reference
thickness of 32mm.

There seems to be some disagreement among researchers as to the actual cause of


thickness effect and how it is to be treated in design. The API RP 2A (1987), for
example, considers that reduction in fatigue life with increased plate thickness can be
4-32
completely avoided if a better weld profile can be achieved and consequently does not
suggest any reduction in fatigue life with plate thickness. The thickness effect has
been a topic of joint European fatigue research programs over the last decade and the
final body of results were published in a number of papers at the 1987 Deift
conference on Steel in Marine Structures, see for example Berge et al (1987). It is
now understood that the following factors could be contributing to this thickness
effect, see for example Berge (1985):

- Stress Gradient Effect: For components subjected to predominantly bending


stresses, the stress gradient will be steeper in a thinner plate compared to a
thicker plate. Consequently, for geometrically similar joints with same surface
stress, the tip of a surface crack grows at a much higher stress in a thicker
plate than a crack of the same depth in a thinner plate. This higher initiation
stress and rapid early crack growth more than compensates for the longer total
length to propagate and results in a lower fatigue life for the thicker plate.

- Weld Geometry Effect: For as-welded joints, the weld toe radius p depends
on the conditions of the last weld pass and unfortunately remains independent
of plate thickness. From Eqn.4.9 it is clear that for geometrically similar joints
with same weld toe radius the notch stress concentration factor will be higher
in a thicker plate than in a thinner plate resulting in lower fatigue life for the
thicker component.

- Metallurgical Effect: A thicker plate in general will have coarser grains,


higher residual stresses and more unfavourably shaped inclusions than the
heavily deformed thinner plate resulting in higher strength for the thinner plate.

- Statistical Effect: For the same surface stress the volume of stressed material
will be larger in a plate of higher thickness resulting in a higher probability of
finding a defect of significant size thereby reducing the fatigue life.
4-33
The weld notch stress concentration can be significantly reduced by providing a
smooth profile at the weld-parent plate intersection. A reduced notch stress will also
reduce the difference in stress gradient between plates of different thickness resulting
in a reduced thickness effect as has been observed in many tests, see for example
Sablok and Hart (1990). The recent AWS code (1988) therefore recommends for
progressively improved weld profiles with increased thickness and a lower S-N curve
for as-welded joints. Improved profiling of welds reduces the weld notch stress and
also the geometric stress by moving the weld toe further away from the brace/chord
intersection resulting in considerable improvement of fatigue life. However, profiling
cannot avoid the metallurgical and statistical effects and a reduced level of stress
gradient effect will still be present. For this reason the approach followed by D.En.
to apply an explicit thickness correction penalty seems to be a better way of treating
the thickness effect in design.

4.4.2.3 Effect of Sea Water Environment:

The basic S-N curve was derived from tests conducted in air but tubular joints in an
offshore structure are subjected to a range of environments such as, spray, free
corrosion under full immersion and cathodic protection. Fatigue life of a joint is
found to reduce for both free corrosion and cathodic protection environments at high
stress levels whilst at low stress levels no detrimental effect is observed. To account
for this effect the fatigue life obtained from the basic S-N curve is reduced by a factor
of 2 in the high-stress, low-cycle (N<10 7) region while the basic S-N curve in air is
used for the high-cycle, low-stress regime. Note that the 1984 D.En. guidelines
allowed the use of basic S-N curve for cathodically protected joints. The new
revisions are derived mainly from tests on plated joints and a few tubular joints.

4.4.2.4 Effect of Weld Improvement Techniques:

Weld improvement techniques such as toe grinding, profiling and post-weld heat
treatment are found to have a beneficial effect on fatigue life. For toe-grinding to be
4-34
effective D.En. (1984) stipulates that the grinding should produce a smooth concave
profile and should extend into the parent plate to about 0.5mm below any visible
undercut at the weld toe. Provided cracking does not initiate from weld root an
increase in fatigue life by a factor of 2 is allowed for joints protected from corrosion.

Weld profiling to produce welds which merge smoothly with the adjoining base metal
should improve the fatigue life of a joint as it reduces the weld notch stress
concentration factor. The API code allows benefit to be gained in fatigue life for
welds with profile control by using X curve instead of the lower X' curve for welds
without such control. However, D.En. does not allow any benefit to be gained by
profile control alone if not accompanied by toe grinding.

Post-weld heat treatment (PWHT) is supposed to remove residual stresses and thereby
increase fatigue life since the effective stress-range is reduced for a fully reversed
loading. Again D.En. guidelines do not allow any benefit to be gained from this since
it is felt that PWHT may not be fully effective in removing all residual stresses in
practical size joints.

4.4.2.5 Effect of Variable Amplitude Loading:

The basic S-N curve was developed from tests on joints subjected to constant
amplitude cyclic loading while the fatigue loading on an offshore structure is of a
variable amplitude type. From constant amplitude tests it is observed that below a
certain level of stress-range, called "threshold stress-range", an indefinitely large
number of cycles can be sustained without any fatigue damage. For an air
environment this is found to be about 53 N/mm2 but is assumed to correspond to
N=107 cycles on the basic S-N curve. However, under variable amplitude loading the
higher stress cycles in the loading spectrum will cause growth of the defect thereby
reducing the value of the threshold stress range (in terms of fracture mechanics, for
a constant threshold stress intensity factor range the threshold stress-range decreases
as the crack size increases). Because of this, as time passes, more and more stress
4-35

cycles which were previously below the threshold level can contribute to fatigue
damage (crack growth). In order to account for this a change in slope from -1/3 to -
1/5 is used in the basic S-N curve which provides a variable threshold level for small
amplitude load cycles. In order to calculate cumulative fatigue damage under variable
amplitude loading a linear summation based on Miner's rule is recommended.

4.4.2.6 Computation of Fatigue Damage:

In an S-N approach fatigue damage is quantified in terms of Miner's damage


summation. Miner's rule assumes that every stress cycle causes some fatigue damage
and that damage caused by various stress cycles are linearly additive. Miner's rule
neglects all stress-cycle interaction effects and thus damage due to n(S) cycles of
constant amplitude loading of stress-range S can be expressed as


D - n(S) (4.13)
N(S)

where N(S) is the total number of stress cycles to failure at the same stress-range S
and D is the damage indicator. From the above it is clear that D should be. equal to
unity for failure. Assuming an S-N curve of constant slope m, it can be expressed as

NSm-K (4.14)

Combining the two equations the damage indicator can be expressed as

D - !n5m (4.15)
"K

For variable amplitude loading, this can be generalised as


(4.16)
- ..g
1m
4-36
In the above, it is assumed that all the corrections for plate thickness, environment,
weld improvement etc. have been incorporated through an appropriate value of K. If
D1 is the damage computed for all stress cycles in one year, then lfDiy gives the
fatigue life or the number of years to failure of the joint.

In Eq.4.16 the summation term represents the cumulative effect of all stress cycles in
a reference period and is referred to as "fatigue loading function". Methods for
response analysis of offshore structures and development of long-term hot-spot
stress-range distribution have already been discussed in chapter 3. The summation
term in Eq.4.16 can be easily calculated using the proposed mixed-Weibull density
function for stress-ranges as shown before. When the S-N curve is bi-linear, as the
T curve given by D.En., the partial expectations of stress cycles falling in the two
regimes of the S-N curves are computed separately, using their respective slopes, and
added together to obtain the cumulative loading term as shown in section 3.5.4.

4.4.2.7 Comments on the use of S-N Approach:

The S-N method of fatigue analysis is very simple to apply and has been widely used
in the offshore industry. Since an S-N curve has been developed from a direct
observation of fatigue lives of a set of joints, it incorporates the effect of all known
and unknown factors on fatigue life which were present in the test data.
Consequently, when used with joint types and loading conditions which are similar to
those from which the S-N curve was developed, a high degree of confidence can be
assigned to predicted fatigue lives. However, in practical applications extrapolations
outside the original data-sets have to be made and in general the method suffers from
the following disadvantages:

(a) Fatigue tests on tubular joints required for developing S-N curves are very
expensive and every additional factor, such as thickness effect, weld toe
grinding etc., that is to be accounted necessitates a multiplicity of new tests.
4-37

(b) S-N curves cannot be easily extended to cases other than those covered by the
original data-base used for developing the curves. Thus a change of material
or major changes in joint configuration and sizes require new S-N curves to
be developed. For example, the current D.En. T curve is not validated for
overlapping and stiffened joints and for materials with yield stress higher than
400 N/mm2 under a corrosive environment.

(c) In developing S-N curves, joints of different types, different weld geometries
and different loading modes are all combined into one data-set and
consequently a very large scatter in fatigue lives is observed over the data-set.
Because of this the predicted fatigue life for a joint using the mean value S-N
curve can be significantly different from the true mean life for that joint type.
The design S-N curve may be used to provide a "safe" (often over
conservative) estimate of fatigue life. Moreover, the influence of such
important factors as weld geometiy, residual stresses, degree of bending etc.
cannot be evaluated explicitly.

(d) Fatigue tests on tubular joints used in developing S-N curves were all of
simple geometry subjected to constant amplitude loading of a single mode (eg.
axial loading). However, in reality tubular joints in offshore structures are
multi-planar, and are subjected to variable amplitude combined loading. For
multi-planar joints in the structure the boundary conditions and stiffening
effects felt by a growing crack may be different from those of simple joints
tested in laboratory. Under variable amplitude loading stress cycle interaction
effects are important, particularly under corrosive environment. The effect of
wave directionality and combined loading is a "wandering hot-spot" which in
general should give longer fatigue lives compared to laboratory tests.
However, this may also result in multiple cracks developing over a much wider
area around the tubular intersection, and if these coalesce early in life can
result in much lower fatigue life than expected. These factors contribute
significant uncertainties in addition to those in estimating fatigue loading.
4-38

(e) During in-service inspection if a crack of significant size is found in a joint,


the S-N method cannot be used to estimate the remaining life of the joint.
Similarly, when extensive information is available about weld defect size and
distribution from fabrication inspection, this information cannot be
meaningfully utilised in an S-N type of analysis.

Because of the above reasons the emphasis of research, in recent years, has turned
towards the use of fracture mechanics methods for the fatigue analysis of tubular
joints and this is considered in the next section.

4.4.3 Fatigue Analysis Using Fracture Mechanics Approach:

In welded joints the presence of defects such as slag inclusions, porosity, undercuts,
lack of fusion etc. are, to a large extent, unavoidable. Fatigue cracks often initiate
from these weld defects and then propagate further to cause failure of the joint. For
tubular joints it is known that most of the life is spent in crack propagation. These
observations suggest the use of fracture mechanics principles for fatigue assessment.

Whenever a crack is observed (or postulated) its propensity to extend under a static
load (fracture) or its sub-critical propagation under cyclic loads can be more rationally
explained using fracture mechanics principles. Fracture mechanics provides an
accurate description of the stress-strain fields around the crack and failure conditions
for the component containing a defect. More details of the basis and applicability of
various fracture mechanics methods can be found in text-books on the subject, see for
example Ewalds and Wanhill (1985). Here it suffices to note that fatigue in tubular
joints involves typically low stress levels and the crack tip plastic zone will be small
compared to crack dimensions and hence 'linear elastic fracture mechanics' will be
applicable. A fracture mechanics based methodology is developed in the following
for fatigue analysis of tubular joints. But first a brief review of the various crack
propagation models and stress intensity factor solutions is given as these constitute the
most important aspects of any fracture mechanics fatigue analysis procedure.
4-39
4.4.3.1 Review of Crack Propagation Models:

Analogous to the S-N approach, the fracture mechanics approach uses the stress
intensity factor range AK to correlate the crack propagation rate da/dn. A crack
propagation model typically gives a functional relation of the form


dn -f(AK) (4.17)

I
The form of the function may be either deterministic or probabilistic and accordingly
the various crack propagation models can be broadly grouped into deterministic and
probabilistic models. Deterministic models can be further categorised into Paris type
or Elber type models as discussed in the following.

Paris Type Models:


Perhaps the most commonly used model for fatigue analysis is the one given by Paris
and Erdogan (1963), commonly known as the Paris law which can be given as

- C(AK) m for AK>O (4.18)


dn

where C and m are material constants which can be determined from experiments on
simple specimens of the material.

A typical plot of da/dn versus AK on a logarithmic scale obtained from experimental


data is shown in Fig.4.7. The crack propagation rate curve exhibits a sigmoidal
behaviour which can be considered to fall in three regimes based on the value of AK
as shown in the figure. It can be seen that the Paris law correlates well with the
observed crack propagation behaviour in regime II where log(da/dn) is more or less
linearly related to log(AK). The crack propagation rate in this regime is typically
observed to be of the order of iO to iO mm/cycle and shows relatively little
influence of material microstructure, mean stress etc. In regime III, when the
maximum stress intensity factor K in the applied stress cycle exceeds a value of
about 70% of the material plane strain fracture toughness K1 c , the crack growth rate
accelerates rapidly leading to unstable fracture at Km=Ki• On the other hand for
4-40

small values of stress intensity factor range in regime I a threshold behaviour is


observed. When AK falls below some threshold value no crack growth occurs. The
crack propagation rate in both regimes I and III is seen to be significantly influenced
by the material microstructure, the environment and stress ratio of the cyclic loading.

WI
-a

£Kth £kthD

Log (H)—.

Fig.4.7: Typical variation of crack propagation rate with stress intensity factor range

In practical applications the observed crack propagation behaviour is often idealised


as shown by the dotted line in Fig.4.7 and the Paris law is assumed valid even in
regimes I and ifi. Various modifications have been suggested to the basic Paris law
to account for the influence of stress ratio in regimes I and III. In regime I, a
correction suggested by Kiesnil and Lucas (1972) and in regime Ill the correction by
Forman ci al (1964) is often used. These corrections can be combined and stated as

da C(AKmAK1l)
- - for AK>AKth (4.19)
dn (1R)(K1cAK)
4-41
However, these corrections are best fits to limited experimental data and fail to
describe crack propagation correctly and consistently in all cases and hence the simple
Paris law should be preferred.

Elber Type Models:


These models use a phenomenon called 'crack closure' discovered by Elber (1971).
It is observed that during the unloading part of a cycle the crack closes before the zero
load is reached (i.e when the load is still tensile). This is explained from a
consideration of crack tip plasticity. At the tip of a growing crack each loading cycle
generates a monotonic plastic zone during loading and a much smaller reversed plastic
zone during unloading (approximately one-quarter of the monotonic plastic zone).
This residual plastic deformation leaves a material with residual tensile strains in the
crack flanks. Thus, during unloading, as the crack closes the reaction of the
surrounding elastic material which is undergoing relaxation causes a reactive force on
the plastic material in crack flanks. Thus at zero load significant residual compressive
stresses will be present on the crack face. Consequently during the loading part of the
cycle the applied load needs to overcome these residual compressive stresses before
the crack can be opened.

Arguing that crack propagation can only occur during the portion of the loading cycle
in which the crack is fully open, Elber defined an 'effective stress intensity factor
range' as (AK) Cfl=Kfl -KO where K0 corresponds to the crack opening stress S, at
which the crack is just fully open. The equations 4.17 and 4.18 can now be used by
substituting aK by AK and by 1K.e ff. When the experimental crack growth
data is re-plotted in terms of da/dn versus K11 a good correlation is obtained in all
three regimes regardless of the stress ratio. However, to use AK one needs to know
the value of K, which has to be obtained from experiments. Since iK is dependent
on R it is to be expected that K will be dependent on R and in general AKff
approaches iK at high R (>0.5) values.
4-42
Probabilistic Crack Propagation Models:
Fatigue crack propagation is seen to be an inherently random phenomenon and
experimental results for Paris law constants often exhibit very high scatter. Virkier
et al (1979) have carried out carefully controlled tests on 68 replicate centre cracked
specimens under constant amplitude loading with initial half crack length of 9 mm.
A plot of a vs n for all the 68 specimens, shown in Fig.4.8a, exhibits considerable
scatter suggesting for the use of fully probabilistic models to describe crack growth.
Several attempts have been made to describe this scatter in experimental results
adequately and to obtain the probability distribution for the crack size a(n) after n
cycles of loading by modelling crack growth as a stochastic process. A number of
stochastic crack growth models have been reported in the literature as reviewed by
Bogdanhoff and Kozin (1984), Palmberg et al (1986), Yao et al (1986) and Madsen
et al (1986). Recently Langley (1989) has compared the performance of various
models against the experimental results of Virkler and others. Some of the important
probabilistic models are summarized below and their relative performance noted. In
general these models express the crack propagation rate as


- X(a,n)t(a,S,n...) (4.20)
dn

where (') is a deterministic function and X(') is a stochastic process. Alternatively,


in terms of a damage indicator ii, which is a function of crack and specimen geometry
alone, this can be expressed as


- (4.21)
dn

where (') is a deterministic function and L() is a stochastic process. Several


models have been used for X(') or L(') to randomize the crack growth law.

Random Variable Model: The simplest case is to assume that L(t,n) is a random
variable such that L '(it,n) has a mean value of unity and a variance of v1 say. Then
Eq.4.21 can be inverted and integrated to obtain the statistics of i. Conversely the
statistics of n, the number of cycles to reach a damage state j, from an initial damage
Vo can also be obtained.

4 43

REP ICPTE R VS. N OTR


6 0 C T5 P P (ST
66PP CRET S
OELTR P : l 2 P
P PX - 5 2 U' P
RC: 900MM
R;.20
I =.
/ ij
,Gx •U=UPU•' IJ
I

(a) coo

iox
a

IC X

oxi -,
a. . I3 ISO
N IYCLE5) X10 S

mm 60

50

40

a
30

70

(b)
30

0
0 50 100 150 200 250 300

)UJfl&
mmo

50

40

a
30

20

(c)
30

0
0 50 300 350 200 250 XC

N/103

Fig.4.8: crack length versus number of stress cycles. (a) Scatter in test results,
Virkier et al (1979) (b) C deterministic (c) C random variable
4-44

In comparison with experimental results, see Langley (1989), this model gives a good
fit depending on the choice of parameters for ('). However, when the initial damage
state is changed the model severely under predicts the variance, Var[n 2-n1], of the
number of cycles required to reach damage state V2(n2) from 1 (n1 ), i.e. when 'Vi#Wo.
Thus its general applicability beyond that covered by the experiments is questionable.

White Noise Markov Model: Lin and Yang (1983) have considered the process L(')
to be of the form 1+L2(n) where L2 (n) is a zero mean Gaussian white noise. In this
case the crack length is a Markov process which allows analytical solutions for the
statistical moments of the number of cycles to reach a particular crack size. The
statistical parameters required to define the process L 2(n) are obtained from simple
experiments. Using the same model Sobzcyk (1986) and Madsen et al (1986) have
shown that for a fixed n, t(a) follows a Gaussian distribution while the number of
cycles n(V) to reach a particular damage state has an inverse Gaussian distribution.
The moments of the inverse Gaussian distribution can be obtained in closed form.
However, Langley (1989) has shown that this model does not give a good fit to the
experimental data at larger crack lengths and severely overestimates Var[n2-n1].

Bogdanoff and Kozin Model: Bogdanoff and Kozin (1985) have proposed modelling
the crack size as a Markov process but discretely valued and discretely parametered,
namely a Markov chain. The method describes the damage accumulated during each
cycle (a storm, earthquake, etc). The transition probabilities of reaching one damage
state from a lower damage state is obtained from simple experiments. With the proper
choice of parameters the mean and variance of n can be predicted accurately.
However, since the Markov model essentially assumes that (n 2-n1) is statistically
independent of n1 , the model is seen to overestimate Var[n 2-n 1], Langley (1989).

Random Process Model: Yang et al and Ortiz (c.f. Madsen et al, 1986) have
considered the process L(qi,n)=L 3(n) as a stationary Lognormal process. In this case,
however, the statistics of have to be obtained by numerical integration. The
distribution of 'qi(a) is not readily obtainable and the distribution of a(n) even more
4-45

complicated to determine. This model lies somewhere between the random variable
model and the Markov model and is expected to give good results. Yang et a! (c.f.
Madsen et al, 1986) have proposed second moment approximations for this model.

Ditlevsen and Olesen (c.f. Yao et al, 1986), on the other hand, treat the scatter in
experimental data to be comprised of a random "between" specimen variation and a
random "within" specimen variation. The findings indicate that the random vector
variation between specimens can be reasonably well described by a joint
normal-lognormal distribution and the within specimen variation is such that the
number of cycles needed to grow a crack by a given length is well described by an
inverse Gaussian distribution.

From the above discussion it is to be noted that the stochastic models require
considerably large experimental data to develop their model parameters. The
distribution for the number of cycles required to grow a crack to a given length n(a)
or, alternatively, crack size after a given number of cycles of loading a(n) is not
available in closed form in most cases and despite their complexity do not always give
good results. In a practical reliability analysis procedure for tubular joints of offshore
structures, where the uncertainties in loading are usually much larger than in crack
growth material properties, a simple deterministic model of the Paris or Elber type
should be preferred. A plot of crack depth verses number of cycles using the Paris
law with deterministic values (j)ossibly mean values) of material constants C and m
would typically lead to a single curve as shown in Fig.4.8b. However, the uncertainty
in crack growth material properties can be treated by considering C or m (or both) as
random variables as presented in chapter 5. Considering C as random, independent
and identically distributed from specimen to specimen, and with m fixed would
typically lead to sample functions as shown in Fig. 4.8c. These curves are however
smooth and no intermingling takes place as observed in Virkler's data set.
4-46
4.4.3.2 Review of Stress Intensity Factor Solutions:

The calculation of stress intensity factors for tubular joint cracks is an extremely
difficult task owing to the complex geometry and the three dimensional nature of the
stress distribution. The fatigue cracks observed in these joints are highly irregular and
warped; often exhibiting mixed mode behaviour. Because of this complexity an exact
analytical solution does not exist. In the past few years a number of attempts have
been made to develop approximate solutions for tubular joint cracks. These can be
broadly grouped under four heads as discussed in the following.

Experimental Methods:
Hibberd and Dover (1977) first proposed a method for determining stress intensity
factors from crack propagation data on tubular joints. Since then this approach has
been followed by a number of workers and a library of crack propagation data and
compliance functions have been compiled for different types of joints subjected to
different loading conditions, see Dharmavasan and Dover (1988). If the Paris law for
crack growth is assumed to govern the crack growth in tubular joints and the material
constants C and m are known from experiments on specimens of the same material,
the crack growth data on tubular joints can be represented as

da m
- - CAKe,q, - CYSm(ita)Mtz
dn
(4.22)
(da/dn)m
CXP
C'"S (E a)'12

where is the 'stress intensity geometry correction factor' or 'compliance function'


obtained from the experiments. The advantage of this method is that it accounts for
the natural crack shape development appropnate for fatigue crack growth in tubular
joints. However, fatigue crack growth experiments on tubular joints are highly
expensive and this approach cannot be used solely for determining stress intensity
factors. Besides, the stress intensity factors are calibrated as a function of crack depth
alone and crack aspect ratio is not explicitly modelled. The developed stress intensity
4-47

factors cannot be used for an observed defect which may have an aspect ratio quite
different from that observed in experiments. Furthermore, the computed values of
Y will be sensitive to the assumed values of C and m and the inherent scatter in
their values will also introduce randomness in the computed stress intensity factors.
However, the compiled crack growth data and compliance functions can be very useful
in comparing the accuracy of any numerical or semi-theoretical approaches.

Numerical Methods:
In recent years, a number of studies using 3-D finite element analysis for calculation
of stress intensity factors have been reported in the literature, see for example Rhee
and Salama (1986), Ritchie (1987) etc. In these studies the crack is explicitly
modelled in the idealization of the cracked component and hence these methods are
called 'direct methods'. From the stress analysis of the cracked body stress intensity
factors can be calculated from the displaced shape of the crack or from the stress field
close to the crack tip. if a fatigue crack growth analysis is to be performed the stress
intensity factors have to be first computed for various assumed crack geometries.
Using 20-node isoparametric brick elements Ritchie (1987) has demonstrated very
good agreement with experimental results. However, the direct methods are extremely
expensive with typical reported solution times of the order of 10 hours on a Vax
computer. Another disadvantage is that for fatigue crack growth analysis the crack
shape development should be known or assumed in advance which makes these
methods less suitable as predictive tools. These methods should preferably be used
to establish benchmarks against which simpler methods of analysis could be checked.

Nicholson (1987) has proposed the use of an elasto-plastic line-spring model for
determining stress intensity factors which was demonstrated on cracked steel tubes.
Huang and Hancock (1988) and Haswell (1990) have extended this method to the
analysis of tubular T joints. The advantage of this method is that it reduces the
complexity of a 3-D problem to a 2-D shell or plate analysis, thus saving enormous
computer time. In this approach a partially penetrated surface crack is represented as
a through-thickness crack with springs attached between the faces. The spring
stiffnesses are calculated to represent the stiffness of the remaining ligament of the
4-48

cracked plate. This method holds considerable promise for the efficient calculation
of stress intensity factors, however, its general applicability to non-uniform 3-D stress
states, such as in tubular joints, needs further validation.

Since the calculation of stress intensity factors using the direct methods described
above for different crack shapes is expensive, a number of 'indirect methods' have
been developed. In these methods, a stress analysis of the un-flawed body is
performed initially, and, maldng use of superposition principles, the stresses acting on
the crack surfaces are used in a separate analysis to compute stress intensity factors.
One of the most powerful indirect method for computing stress intensity factors for
general stress fields is the 'weight function method'. If a reference stress intensity
factor and the associated crack opening displacement field normal to the crack face
is known for a simple stress system, a weight function may be calculated by the
method suggested by Bueckner and Rice, see Mattheck et al (1983). A weight
function is seen to be a property of the geometry of the body and the crack. Once a
weight function is available for cracked a component, the stress intensity factors for
any other stress field can be calculated by integrating the stress field, factored by the
weight function, over the crack face. This method has been successfully used for
calculating stress intensity factors for plates under stress gradients, see Mattheck et al
(1983). However, general weight functions for 3-D tubular joint geometries are not
yet available and will have to be first derived using full 3-D finite element analyses.
'Boundary integral equation' methods have also been used sometimes with significant
advantages in deriving weight functions, see for example Heliot et al (1980) in which
this method is used for deriving weight functions and stress intensity factors for
semi-elliptical cracks in cylinders subjected to stress gradients.

Empirical Methods:

If crack propagation data is available from a sufficiently large number of tubular joint
fatigue tests for which experimental compliance functions are calculated by the
method described before using Eq.4.22, then simple empirical expressions can be fitted
for Y,. Dover and Dharmavasan (1982) used this approach and proposed an
'average stress model' in which the compliance function is correlated in terms of the
4-49

average stress around the tubular intersection. This model was later modified by Kam
et al (1987) who proposed a 'two phase model' which is similarly based on the
average stress and in addition joint geometry parameter (3 and the plate thickness.
This model was based on an experimental data-set of two T, two Y and three X joints.
The model is shown to give good correlation with crack growth results from similar
tests. However, the limitations of these models are that stress intensity factors are
expressed as a function of only the crack depth and hence these models cannot be
used to predict fatigue crack growth in the surface direction. Moreover, since typical
weld geometry, residual stress, etc. have been implicitly included in the model the
sensitivity of fatigue life to these important parameters cannot be studied using these
models. Since the data-set used in deriving these models is fairly small their
applicability to other joint geometries, crack shapes and loading needs to be validated.

Method of Superposition:
A number of stress intensity factor solutions are now available for idealised
components and crack geometries and simple stress distributions and these have been
compiled into hand-books such as those by Tada et a! (1973) and Rooke and
Cartwright (1976). Normally these solutions are expressed as correction factors for
the classical solution of K=YrI(ita) with Y=1 for a through-thickness centre crack in
an infinite plate subjected to uniform stress s remote from the crack. For a given
geometry the stress intensity factors can then be determined by superposing these
standard solutions to apply corrections for one or more of the features relating to the
crack geometry, stress field etc of the component under consideration. This approach
can be conveniently used to determine stress intensity factors for tubular joints. A
number of studies on tubular joints, see for example Delft et al (1986), have used the
stress intensity factors proposed by Newman and Raju (1981) or those by Scott and
Thorpe (1981). These solutions are derived for semi-elliptical surface cracks in
un-welded flat plates subjected to uniform stress across the width of the plate.
However, the tubular joint cracks are subjected to a varying stress field both in the
through-thickness direction and along the tubular intersection. The effect of weld
geometry is very important in the initial phase of crack growth in tubular joints which
4-50

is also characterised by multiple crack initiation and coalescence. The boundary


conditions used in developing flat plate solutions and the local stiffness changes which
a growing crack experiences in a tubular joint can be quite different from those in flat
plates. For these reasons classical flat plate solutions should not be used for tubular
joints without accounting for the additional characteristics of the tubular joint problem
as shown in the proposed methodology for fracture mechanics fatigue analysis of
tubular joints presented in the next section.

4.4.4 Proposed Methodology for Fatigue Analysis of Tubular Joints:

From the foregoing review it is clear that although the fracture mechanics method
offers a rigorous framework for the treatment of fatigue process in welded joints, its
application as a design method for tubular joints has not yet been recommended by
standard codes owing mainly to: (a) the lack of simplified but accurate solutions for
stress intensity factors and the enormous computational effort involved in the use of
finite element solutions, (b) the complexity of fatigue crack growth in tubular joints
which is influenced by a number of factors which cannot be easily quantified and
(c) the general complexity of the method compared to the S-N approach. However,
a fracture mechanics approach is almost a necessity in the assessment of significance
of weld defects observed during fabrication or fatigue cracks discovered during
in-service inspections. The fracture mechanics approach is also very useful in a
fatigue reliability assessment of tubular joints wherein the sensitivity of the reliability
of a joint to various uncertain input parameters can be studied. The method offers the
only rational way to treat fatigue deterioration in a system reliability assessment of
offshore structures. A simplified fracture mechanics methodology for fatigue analysis
of tubular joints is developed in the following for use in such situations.
4-51
4.4.4.1 Stress Intensity Factors:

From the review of stress intensity factor solutions in section 4.4.3.2, it is clear that
while the numerical methods and experimental methods are prohibitively expensive
for practical applications, the available empirical methods are limited in their range
of applicability. In the proposed method, therefore, a simplified approach, based on
the method of superposition, is adopted in obtaining stress intensity factors. Starting
from available solutions for a semi-elliptical surface crack in a flat plate, corrections
are applied to account for one or more of the characteristics of the tubular joint
problem as discussed below.

Stress Intensit y Factors for a Semi-Elliptical Crack:


From experiments on tubular joints it is observed that micro-cracks initiate from
surface-breaking defects at the toe of the weld. These micro-cracks coalesce together
to form a single, dominant fatigue crack of roughly semi-elliptical shape, which in
most cases propagates into the chord plate. Thus, as a first approximation, a fatigue
crack in a tubular joint can be idealized as a semi-elliptical surface crack in a flat
plate. Empirical stress intensity factor solutions for a semi-elliptical surface crack in
a plate subjected to membrane and bending stresses through the thickness of the plate
have been developed by Newman and Raju (1981) based on results from 3-D finite
element analysis. For a semi-elliptical crack of depth a and semi-length c, the stress
intensity factor K can be expressed in the form

K - (Hp+l-p)Sg Yg (7taIq) 12 (4.23)

where S is the geometric stress at the crack site, p is the degree-of-bending factor for
the stress distribution through the plate thickness and q is the elliptical integral of the
second kind. The factors H and Yg are functions of the relative crack depth all', crack
aspect ratio a/c, relative crack length c/W and angle $ to the point on the crack front,
for which expressions have been given in the above reference. In the above T is the
plate thickness and W is the plate width, taken here as the circumference of the brace
member. Using the above equation the stress intensity factor at any point on the crack
front can be determined but the interest here is only on the deepest point of the crack
4-52
in the thickness direction (4=90) and the surface tips (4=O°). Similar solutions have
also been given by Scott and Thorpe (1981). The Newman-Raju solution, in effect,
accounts for corrections due to the free surface, finite geometry and crack elipticity
To account for other features of tubular joints additional corrections are needed.

For using the Newman-Raju solution given above, a knowledge of the total geometric
stress at the crack site and its distribution in the thickness direction is required. To
be commensurate with the simplicity of the proposed model it is suggested that
parametric equations for geometric stress distributions presented in section 4.3.1 may
be used. This will also be convenient in a fully automated fatigue analysis of a large
number of joints in a structure which is often the need in system reliability
assessments. However, when higher accuracy is desired results from an earlier finite
element analysis can be used with polynomial interpolations in a reliability analysis.

Effect of Weld Geometry:


The presence of a weld at the plate surface gives rise to a non-linear notch stress
distribution as shown in Fig.4.6. To account for this notch stress a weld toe correction
factor is used. The calculation of stress intensity factors for a semi-elliptical crack
under a varying stress field is generally very difficult. If the notch stress distribution
along the crack propagation path is known the weld toe correction factor can be
calculated fairly efficiently by the method given by Albrecht and Yamada (1977).
This method is known to be reasonably accurate for edge cracks at weld toes but its
validity for semi-elliptical cracks is not known. The effect of weld geometry on stress
intensity factors can be calculated very accurately using the weight functions given by
Niu and Glinka (1987) for edge cracks and Niu and Glinka (1989) for semi-elliptical
cracks in T butt weidments. The method for semi-elliptical cracks uses a combination
of the weight function for an edge crack at a weld toe with the weight function given
by Mattheck et al (1983) for semi-elliptical cracks in smooth plates to determine stress
intensity factors for semi-elliptical cracks at weld toes. The notch and geometric
stress distribution is integrated numerically over the crack surface.
4-53

In order to use the methods mentioned above, an accurate estimate of notch stress
distribution through the thickness of the plate and around the tubular intersection is
required which can be obtained, for example using a finite element method. However,
when the geometric and/or notch stress distribution is obtained only approximately
these methods do not offer any additional advantage. For practical applications,
simplified parametric expressions have been developed which have a general form of

- ía l w 1a Pw (4.24)
w

where l is the weld leg length, is the weld toe angle and P is the weld toe
radius. Empirical expressions for function Y 3 as a function of relative crack depth and
relative weld leg length for =45 and p,=O are given in BS PD6493 (1989)
separately for membrane and bending components of the through thickness geometric
stress distribution. Similar expressions for functions Y 2 and Y3 can be obtained from
Dijkstra et al (1990). However, it is to be noted that these expressions are developed
from parametric studies using 2-D finite element analysis on edge cracks in weidments
and can be conservative for semi-elliptical cracks, Dijkstra et al (1990). Similar
solutions are not available for semi-elliptical cracks at present.

Effect of Stress Gradient:


In a tubular joint, because of the stress concentration, the geometric stress decreases
rapidly away from the hot-spot. The stress gradient is steeper for nominal bending
moments than for axial load in the brace. Although the hot-spot stress alone could be
used to characterize crack propagation when the crack is small, the variation in crack
face loading should be considered when the crack becomes long and grows into
regions away from the hot-spot. The stress intensity factors calculated from
Newman-Raju solutions are valid for a plate subjected to uniform stress across the
width of the plate. An additional correction is required to account for stress gradient.

Mattheck et al (1983) give a method for calculating stress intensity factors for
semi-elliptical cracks in plates subjected to a varying stress field using the weight
function technique. The Newman-Raju solution is used to calculate the reference
4-54

stress intensity factors and an approximate expression is developed for the reference
crack opening displacements. The required stress intensity factors are then calculated
by using the weight function and integrating numerically the varying stress field over
the crack face. But the marginal accuracy which could be gained from this rigorous
approach will be diluted if the stress distribution around the intersection is modelled
only approximately using empirical equations. To maintain the simplicity of the stress
intensity factor model it is therefore suggested that the effect of stress gradient is
modelled by an 'average stress factor' defined as

= average stress over the crack length


S
(4.25)
geometric hot-spot stress

This stress gradient correction is applied for the crack growth at the surface tips only
and is conservatively neglected for the growth at the deepest point of the crack.

Effect of Load Sheddin g During Crack Growth:


in a tubular joint, the rigid connection between the brace and chord members and
variable stiffness of the chord plate around the tubular intersection gives rise to
geometric stress concentrations at the joint. The through-thickness distribution of
geometric stress therefore comprises a membrane component, which balances the load
transferred through the brace, and a significant bending component, which maintains
displacement compatibility at the joint. Thus if the welded connection is replaced by
a hinge the joint will be able to transfer the same load with only a membrane stress
at the joint and no bending stress.

The introduction of a crack at the tubular intersection reduces the rigid connection
between the brace and the chord locally with a consequent reduction of the stress
concentration at the crack site. Strain gauge monitoring during fatigue tests has shown
that the hot-spot strain falls continuously as the crack propagates through ihe
thickness. This reduction in hot spot stress is expected to be due to the relaxation of
bending stress alone. This load shedding is not modelled in the Newman-Raju
solutions which are developed for flat plates with no stress concentration. An
additional correction, therefore, becomes necessary when using these solutions for
4.55
tubular joints. An explicit load shedding is not required when using a 3-D finite
element model of the tubular joint as stiffness changes at the hot-spot at varying crack
sizes will be automatically updated.

Adequate experimental data are not available at present to quantify the reduction in
hot-spot stress due to crack growth. It is tentatively proposed to model this by
linearly reducing the bending component of the geometric hot-spot stress from its
initial value at the surface to zero as the crack ruptures the chord wall. This
correction is applied for calculating stress intensity factors at the deepest point of the
crack only and no change is made for the stress intensity factor at the surface tip. A
linear moment release has also been suggested by Aaghaalcouchak et al (1989) which
is explained on the basis of different boundary conditions for the flat plate and tubular
joint configurations. The bending component of the geometric stress range at any
crack depth, Sb(a) can be expressed as

Sb( a ) - SbO E(a)


(4.26)
(a) -

where S is the initial value of bending stress range and (a) is the 'bending stress
relaxation factor'.

In addition to the above mentioned corrections it may also be necessary to apply a


correction for the shell action in a tubular joint. However this correction is known to
be small and can be neglected. The final expression for stress intensity factor
considering all the above corrections is given in section 4.4.4.3.

4.4.4.2 Crack Propagation Model:

The proposed crack propagation model is based on a deterministic crack growth law
of the Paris type. Modifications are made to account for the influence of a number
of factors which influence fatigue in offshore structures. Additional changes are made
4-56

to accommodate the special features of crack growth in tubular joints. Using this
model an attempt is made to explain the observed crack growth behaviour in tubular
joints in terms of the important factors which influence crack growth behaviour at
various stages and how these can be modelled using a fracture mechanics approach.

Basic Crack Growth Law:


The propagation of a semi-elliptical crack in a tubular joint is considered in terms of
crack growth in two principle directions viz, its extension in the thickness direction
at the deepest point and a simultaneous extension at the two surface tips.
Corresponding to these directions two crack growth equations can be considered as

dn
(4.27)

dn

where the range of stress intensity factors iKa and AK can be determined as
discussed in the previous section. The material properties are assumed to be same in
both directions. As noted before, this crack growth law models crack propagation
reasonably well in the regime of intermediate stress intensity factor ranges.

The fatigue behaviour in offshore structures is influenced by a number of factors as


summarized in section 4.4.1.1. Using fracture mechanics principles these factors can
be explicitly accounted for in a fatigue assessment procedure. The influence of joint
geometry is reflected in the determination of geometric stress distribution at a joint as
explained in section 4.3.1. The weld geometry influences notch stress distribution the
influence of which on stress intensity factors is accounted through a weld geometry
correction factor as given in the previous section. The influence of remaining factors
is considered in the following.
4-57
EfTect of Stress Ratio:
For each loading cycle a stress ratio can be defined as

R-
'I
SrIlax
(4.28)
Kmj,,
R
K

where R.g is called a 'nominal stress ratio' which is calculated from the minimum and
maximum values of the applied cyclic stress alone while R is a 'stress intensity factor
ratio' (simply called stress ratio) is based on the total stress which includes applied
cyclic stress, applied static stress, residual stress etc. The influence of the stress ratio
can be seen from Fig.4.7 which shows that crack propagation rate increases as the R
ratio becomes more positive. However, the influence of R is more pronounced in
regimes I and ifi and is highly dependent on the microstructural properties of the
material. The effect of stress ratio in these regimes can be accounted through the
modifications suggested by Kiesnil and Lucas (1972) and Forman (1964) as given by
Eq.4.19. In structural steels the crack propagation in regime II is found to be almost
independent of stress ratio. In addition to this influence on crack propagation rate, the
stress ratio also influences threshold stress intensity factor range and effective stress
range as discussed in the following.

Effect of Residual Stresses:


The effect of tensile residual stresses is to alter the stress ratio and stress intensity
factor range experienced at the crack tip during load cycling which are calculated as

K-K
mui +Kre:
app .mui

K
mat -Kapp .maz res

AK -K
mat-K
mui -K -K
app .mat app.mut for K,,>0
(4.29)
AK - K,, - Kapp ,,+ Kres for K^0
R - K j,,/K for K>O
R - 0 for K^0
4-58

Here it is implicitly assumed that for K 7J ^0 (i.e. under net compressive stress) the
crack will be closed and cannot propagate. Only that part of the loading cycle for
which the crack remains open is considered to be effective in crack propagation.
From the above equations it can be seen that residual stresses do not have any
influence on K as long as K>O. For K^O, i.e. when the applied stress cycle has
a very high compressive component, residual stresses will have a significant influence.
However, the stress ratio is altered in both cases which in turn has an influence on
crack propagation as mentioned above. The presence of any other stresses such as due
to static loading, buoyancy etc which contribute to the mean stress will have a similar
influence and should be considered in calculating iNK.

The calculation of K should be based on the through-thickness distribution of


residual stresses which is idealized as a combination of bending and membrane
components as in Fig.4.6. The residual stresses are believed to dissipate as the crack
propagates through the thickness, the exact nature of which is however not known.
This is modelled here by linearly reducing the bending component of residual stress
from its initial value at the surface to zero as the crack ruptures the plate wall. Thus,
effective residual stress for any crack depth, S(a) can be expressed as

Sr (a) — Sb ,(a) + S in res

- [p (a) + 1 - p } Sres0 (4.30)

(a) — 1-v)

where S 0 is the initial value of residual stress at the surface, p is the degree of
bending factor and (a) is a 'residual stress relaxation' factor. This relaxation is
assumed to be effective for the deepest point of the crack only and no relaxation is
used at the surface tips.

The effect of residual stresses and stress ratio R can be conveniently introduced into
the crack growth law through an effective stress intensity factor ratio U defined as
4-59

- AK K,_K0
(4.31a)
AK K - K

where K, is the crack opening stress intensity factor which is normally regarded as
a material property to be determined from experiments. But here K is introduced
only as a mathematical convenience. This assumption reduces the crack growth law
to a Paris type with AK in Eq.4.27 now replaced by U AK. In calculating U from the
above equation K, is taken as zero for K. <0 (i.e R<0) and K,=K for K>O
(i.e R>0). This helps AK to be determined on the basis of nominal cyclic stress range
alone as AK=YS/(7ta). Thus the effect of residual stresses and mean stress is
introduced through the factor U. Noting that the
expression for U can be modified and written in terms of R as

K,, - Kmjn - (K0,, - Kmi,i)


U-
K Km

- 1 -(1/AK) [K0,, Kmu,]


(4.31b)
- 1[RI(l_R)][(Kop/Kmijpi)_1]
-1 for R>O
- 1/(l-R) for R<0

Effect of Fati2ue Threshold:


From Fig.4.7 it can be seen that when the applied stress intensity factor range AK falls
below a threshold value, called the 'threshold stress intensity factor range AK,', crack
propagation does not occur. The exact mechanism of fatigue threshold is not
completely understood but it is regarded as a material property. Threshold value can
be determined from experiments in which the applied AK is gradually reduced until
the crack propagation stops. The threshold value is seen to be significantly influenced
by material micro-structure, environment, stress ratio and the value of K, used
during experiments, Castro et al (1987). The influence of stress ratio on threshold
stress intensity factor range can be explicitly modelled as suggested by BS PD6493
(1989) using
4-60

th - fR tihO
(4.32)
fR - (190- 144R)/190 - 1-0.757R

where AK is the threshold value determined from experiments at R=0.

Effect of Sea Water Environment:


Fatigue crack growth under corrosive conditions exhibits a highly complex behaviour
which is not yet fully understood. A number of tests were carried out on welded plate
specimens under sea water, with or without cathodic protection during the UKOSRP
project, see D.En. (1987, 0Th 265). It is now regarded that fatigue under corrosive
conditions is a combination of three mechanisms each of which may be dominant at
different ranges of stress intensity, see Solin (1990). These mechanisms are: (1) metal
fatigue in which crack propagation rate may be similar to that in air, (2) stress
corrosion fatigue during which crack growth may be relatively independent of applied
stress intensity factor range resulting in a plateau in the da/dn vs AK curve, and
(3) true corrosion fatigue in which there is an interaction of corrosion and fatigue
processes resulting in accelerated crack propagation rate. A three stage da/dn vs AK
curve is often considered, as shown in Fig.4.9 to model fatigue crack growth
behaviour in corrosive conditions.

Lc1(d/dn)

I I I
£k, £1C4

S B1, B, 84

Fig.4.9: Schematic three-stage model for corrosion fatigue crack growth


4-61

However, the limits defming the three segments and their slopes may change
considerably depending on such factors as material micro-structure, test temperature,
oxygen content, cathodic protection, loading frequency, stress ratio, constant vs
variable amplitude loading etc.

It is not the intention here to develop a suitable model to describe crack growth data
under corrosion. If a three stage model as shown in Fig.4.9 is valid, the overall
methodology described in this section can still be used except however, the influence
of temperature changes, time and stress cycle interaction which may occur in a real
structure cannot be explicitly modelled. As a simplification the Paris model can be
used with modified values of parameters C and m to account for corrosion fatigue.

Effect of Variable Amplitude Loading:


All the expressions presented above are strictly valid for constant amplitude loading
but can be easily extended to the case of variable amplitude loading. Under random
loading, the sequential order of stress cycles is believed to have some influence on
crack propagation rates. A number of approaches for treating variable amplitude
loading have been suggested in the literature which can be broadly grouped into two
categories. The first of these are called load interaction approaches in which the crack
propagation is determined on a cycle by cycle basis considering the effect of one
stress cycle on the subsequent cycles. Of the several interaction models proposed, the
Willenborg model and Wheeler model (c.f. U.E.G, 1985), based on crack tip plasticity
and the de-Koning (1981) model based on crack closure are considered important.
However, these models cannot be easily used in the fatigue assessment of offshore
tubular joints because of the random nature of wave forces on these structures and the
general lack of information on the sequential occurrence of various sea states.
Furthermore, a cycle by cycle damage computation considering interaction needs
time-history simulation and rain-flow counting to determine stress range magnitudes
of each stress cycle in a stress process which is considered to be very expensive in an
iterative reliability analysis of a large number of tubular joints.
4-62
The second category of approaches are called non-load interaction approaches which
effectively compute damage due to each stress cycle using a constant amplitude crack
growth law and sum up the damages without considering any interaction between
stress cycles. Alternatively, an equivalent constant amplitude stress range can be
computed from the knowledge of the probability density function of stress ranges as
explained in section 3.5 of the previous chapter. Thus, considering the basic crack
growth law for the crack propagation in the thickness direction Eq.4.27, for variable
amplitude loading, separating the variables and integrating leads to

a(s) n(i)
fdx (4.33)
-Es1m
Cy(x)[x]m/2

In the above the L.H.S represents the cumulative fatigue damage in growing a crack
from its initial size a0 to a size a(t) after time duration t due to n(t) cycles of loading.
The R.H.S term represents the cumulative effect of n(t) stress cycles of variable
amplitude loading during time t. Methods for developing long-term stress-range
distribution and computing the cumulative loading term has already been presented in
chapter 3. In computing the long-term cumulative fatigue load effect the
non-damaging stress cycles, for which stress intensity factor range is less than the
threshold value, can be filtered out using a threshold factor G(a) defined as, see
Wirsching et al (1987)

E[S m]00
G(a) - _______ (4.34)
E[S]

where s,(a) is the threshold stress-range at a given crack depth calculated from the
threshold stress intensity factor range iK.11. The partial expectations can be calculated
using the density function of stress ranges as given in section 3.5.4 of chapter 3. The
factor 0(a), being a function of crack depth can be moved to the L.H.S of Eq.4.33 and
integrated over crack depth.

In the case of sea water environment where a three stage model is used to describe
crack propagation the computation of cumulative loading is slightly complicated. In
4-63

this case, corresponding to the four terminal points of the three linear segments four
stress range values can be defined as shown in Fig.4.9. Then, using the long-term
density function of stress-ranges the partial expectation of stress cycles occurring in
each of the three segments can be computed separately and added together, see Kam
and Dover (1989).

In addition to the above, certain other factors have to be considered in order to


account for the specific characteristics of crack growth in tubular joints and this is
discussed in the following. From the presentation in section 4.4.1.2, the crack growth
in tubular joints, under laboratory tests, can be considered to fall in three main phases.
These are, Phase-I: This is the initial phase in which a number of cracks develop from
weld toe defects and then coalesce together to form a single dominant crack of
roughly semi-elliptical shape; Phase-fl: Propagation of the semi-elliptical crack in
both the through-thickness direction at its deepest point and along the welded
intersection at its surface tips, until the crack penetrates the plate wall at its deepest
point; and finally, Phase-ifi: Growth of the through-crack along the intersection until
the 'failure' of the joint marked by substantial reduction in joint stiffness. In each of
these three phases different factors have a dominating influence on crack growth
which need to be understood and modelled correctly. Additional modifications are
made to the basic crack growth law to account for special features of the tubular joint
problem as discussed in the following, see also Shetty and Baker (1990b).

Phase-I Crack Growth in Tubular Joints:


In welded tubular joints fatigue cracks are found to originate from weld toe defects
which are to some extent unavoidable. For this reason an initiation (nucleation) period
is often considered to be absent and fatigue life is regarded to comprise only crack
propagation from an initial defect/s. However, the size of initial defects in welds of
sound quality is expected to be very small and these defects initially grow within a
notch zone at the weld toe. The crack growth model should, therefore, be able to
explain the growth of such embryo cracks which later coalesce together to form a
single, dominant crack. In this phase of crack growth, therefore, three important
aspects have to be considered as below.
4-64

Initial Weld Defect Size: Since fatigue cracks originate from weld defects, the size
and distribution of these defects has an important influence on the initial phase of
crack growth. Weld defects, such as undercuts, slag inclusions etc., which occur at
the toe of a weld in regions of high stress concentration are seen more likely to
develop into fatigue cracks. When inspection data are available for a particular joint
the size of initial defect/s will be known and these can be used as inputs to the crack
propagation model. In such cases, the defects may not be strictly semi-elliptical in
shape and a defect schematization convention has to be used, such as the one given
by BS PD6493 (1989), to re-characterize these defects as semi-elliptical surface cracks
for numerical analysis. However, when no information is available on initial weld
defects and the crack propagation model is used as a design/prediction tool, an initial
defect size has to be assumed and the computed fatigue life is seen to be higly
sensitive to the choice of initial defect size. Typical defect sizes at weld toes for
"sound" quality welds is found from microscopic observations to be in the range of
0.1mm to 0.5mm. For reliability assessments, the probability distribution of initial
defect size and occurrence rate should be based on statistical analysis of defect data
collected from structures during fabrication inspection.

Growth of an Embryo Crack: The nominal stress in the brace member is greatly
amplified at the hot-spot location because of the stress concentration effect of the
geometry of the joint and the weld geometry. Hence, even for a stress cycle of
moderate amplitude, the peak stress at the hot-spot may go beyond yield value and
thus, close to the weld toe, there will always be a small region of fully yielded
material, called a 'weld notch plastic zone'. When the weld defect is very small,
which it usually is, and lies within the notch plastic zone, growth of this small crack
cannot be predicted very well using the linear elastic (LEFM) stress intensity factor
range AK and the Paris law of crack growth using Eq.4.27. Experiments by
El Haddad et al (1981) indicate that within a notch plastic zone the crack growth rate
will be much higher than LEFM rates predicted by Eq.4.27. Furthermore, crack
growth rate is found to decrease with increase in crack size (and nominal AK), which
is contrary to LEFM behaviour. Cracks may even become non-propagating as
illustrated in Fig.4.10.
4-65

1. Long crack trend (LEFM)


2. Short crack behaviour
3. Non- propagating crack

Tronsition to
LEFM trend 1

do
dN

'I
A
I"
I %'
I

AK -p

Fig.4.1O: Crack propagation behaviour within a plastic zone, El Haddad et a! (1981)

Smin: 180.0 MPO

100.0 PLostic zone 512e = 2mm


-: ExperimentaL curve
---= Idealization

0.0
- -
i

api

1.0 2.0
Crock length (c) (mm)

Fig.4.11: Variation of crack opening stress within a plastic zone, Akiniwa ci al (1988)
4-66
As the crack grows out of the notch plastic zone the behaviour slowly merges with
the long-crack LEFM behaviour.

This behaviour can be explained using crack closure concepts, see Elber model in
section 4.4.3. If the normal fatigue crack growth behaviour is assumed to be
controlled by crack closure effects, arising out of the constraint of the surrounding
elastic material on a small layer of residual plastic material in the wake of the crack,
this effect no longer takes place for a small crack within a plastic zone, and thus a
lower crack opening stress intensity K0 can be expected. Recent experiments by
Akiniwa et al (1988) show that, within a notch plastic zone, the crack tip will be fully
open even under compressive load and for very short cracks the opening load K can
be equal to the minimum stress intensity of the cycle K.. As the crack moves away
from the notch, the opening load increases and tends to a stabilized value which is
typical of long-crack behaviour as shown in Fig.4.l 1. This variation in K is
idealized here, for simplicity, by the dotted line shown in the figure. Thus

K0,, - [1 - (a /a) ] K,,, for a


(4.35)

where is the weld notch plastic zone size. In order to account for the variation of
K0 in calculating an effective stress intensity factor range, a factor is defined
within the notch zone as

_____
AKeff -______
K,, - K0,,
pz - AK (4.36)

- 1+(U-1)(a/a,,2)

where U is the effective stress intensity factor ratio outside the notch plastic zone.
The effective stress intensity factor range AK=U AK is now substituted into
Eq 4.27 in place of AK.

The above formulation requires the estimation of the notch plastic zone size a. This
can be accurately determined, for a given geometric hot-spot stress, from Eq.4.10 by
4-67

substituting the yield stress F for S(x) and iteratively solving for x which gives the
notch plastic zone size. Under variable amplitude loading, when the magnitude of
individual stress cycles is not known, can be conservatively calculated for the
maximum stress cycle. Alternatively, a constant value of a 1, =3mm may be assumed.

Crack Coalescence: During fatigue tests on tubular joints it has been observed that
multiple cracks initiate in the region of the theoretical hot-spot. These cracks grow
independently for some time and then coalesce to form a single long dominant crack.
The crack growth rate in other cracks which do not coalesce is seen to slow down
once a dominant fatigue crack was formed, Dover et al (1988). Crack coalescence is
thus important for predicting surface crack development. Theoretical models to
account for coalescence effect are, however, very limited. Snijder et al (1988)
presented a coalescence model but this requires, a priori, the information on the
number of cracks initiated and the mean distance between these cracks. This model
cannot be easily applied to crack growth prediction in a real structure where the
stresses are constantly changing due to combined brace loading and a wandering
hot-spot. For simplicity, it is proposed to model the effect of crack coalescence by
considering a damage zone of length equal to 20% of the brace diameter, within which
cracks initiate and grow. Crack coalescence is assumed to be complete when the
leading crack at the hot-spot reaches a depth of 3mm as observed in most fatigue tests.
On coalescence the surface crack length is set equal to the length of the damage zone.
Thus at the end of Phase-I, a semi-elliptical crack of dimension a=3mm and
semi-length c=O.Old is formed. This phase is generally found to occupy about 20%
of the total fatigue life. Weld geometry and residual stresses are found to have
dominating influence on crack growth in Phase-I. While weld root radius is important
for crack initiation its further growth is seen to be controlled by weld toe angle.

Phase-il Crack Growth in Tubular Joints:


The semi-elliptical surface crack formed by coalescence begins to propagate both in
the through-thickness direction and along the welded intersection at the surface. The
crack is now most likely out of the weld notch plastic zone and its propagation can
be well correlated using LEFM methods based on Paris law. For convenience of
4-68

discussion the crack growth in the thickness direction is considered first and surface
crack development is discussed next.

The weld notch stresses have a significant influence in the early stages of crack
growth in the through-thickness direction and this is modelled using the weld
geometry correction factor Y,,. However, the influence of notch stresses drops rapidly
and they are found to have no influence after the crack depth reaches a value of about
a=O.35T. While the Phase-I crack growth is more influenced by the surface value of
residual stresses, in Phase-il the level and distribution of residual stresses is more
important. Furthermore, residual stresses are expected to dissipate as the crack grows
longer and deeper. Unfortunately, the through-thickness distribution of residual
stresses and their dissipation with crack growth is not well researched for tubular
joints. Here, it is proposed to model the effect of residual stresses using the effective
stress intensity factor ratio U and the residual stress relaxation factor as explained
before. One additional factor which has a dominating influence in this phase of crack
growth is 'load shedding'. As the crack grows longer and deeper, it is expected to
reduce the local bending stiffness of the chord plate thereby releasing some of the
stress concentration. This phenomenon is not well understood and less well
quantified. This effect is proposed to be accounted through a bending stress relaxation
factor in the calculation of stress intensity factors as discussed in section 4.4.4.1.

As the crack propagates in the through-thickness direction, there will be a


simultaneous increase in crack length due to crack propagation at its two tips near the
surface. The surface crack growth will continue to be influenced by the weld notch
stress for the entire period of crack growth and this is accounted for by calculating the
weld geometry correction factor for a constant crack depth of a=1.5mm. The
surface crack growth is expected to take place within a notch plastic zone for most of
the life. The relaxation of residual stresses and bending stress can be assumed to have
no influence on the surface crack growth
4-69
Phase-ifi Crack Growth in Tubular Joints:
When the crack penetrates the chord wall at its deepest point, the remaining ligament
of the crack between the surface tips and the deepest point breaks off quickly and the
crack becomes a rectangular through-crack. This crack propagates further along the
weld until 'failure' of the joint occurs. The stress intensity factors for a through crack
in a plate can be calculated from the standard solutions available from the handbooks,
see for example Rooke and Cartwright (1976). However, the crack propagation rate
is expected to vary across the depth of the crack for thicker plates because of change
in stress state. An important factor which influences crack growth in this phase is the
stiffness deterioration of the joint with increasing crack length. This gradual loss in
joint stiffness and consequent reduction in applied stress will be more difficult to
quantify in a real structure and hence this phase of crack growth is usually not
considered by many researchers. Experiments show that this phase of crack growth
may occupy up to 20-30% of the total fatigue life of a joint.

Definition of Fati2ue Failure for Tubular Joints:


The fatigue failure event is not clearly marked in tubular joints. From fatigue tests
on tubular joints it can be seen that after rupture of the plate wall, the crack continues
to propagate along the welded intersection for significant length of time although with
a continuous loss in joint stiffness. In many fatigue tests the loss in stiffness was so
excessive that testing had to be stopped as the desired load could not be maintained.
In tests which could be continued the crack was found to branch out into the chord
at many places and only in few cases that complete separation of the brace from the
chord occurred. Because of this, for a consistent defmition of the failure event,
through-thickness cracking is generally regarded as the failure of the joint. This is
also the definition of failure considered in developing D.En.(1984) design S-N curves.
Although this definition is conservative in a system reliability assessment,
through-thickness cracking is considered here to constitute fatigue failure of a joint.
However, depending on the applied load magnitude, it is possible that a joint fails by
crack instability or fracture before the fatigue crack penetrates the chord wall. This
is considered as a separate limit-state and is the subject of discussion in section 4.5.
4-70

4.4.4.3 Computation of Fatigue Damage:

Based on the crack propagation model and the stress intensity factor solution presented
above, a procedure for computation of fatigue damage is summarized here. For the
sake of simplicity the following derivation assumes variable amplitude fatigue under
air environment. Fatigue damage at any instant, after an exposure period of time t,
or an average number of stress cycles n(t), can be characterised in terms of the crack
dimensions a(t) and c(t). Corresponding to the two principal directions, the basic
crack growth law (Eq.4.27) can be modified and written as

- CM(eG(a)
(4.37)
- CEtKG(c)

where G(a) and G(c) are the threshold factors for the deepest point and the surface tip
to be calculated based on Eq.4.34. The effective stress intensity factor ranges are


effa - U0SY0[Ea/q]1'2
(4.38)
eff.c - USY[1ta/q]''2

where Ua and U are the effective stress intensity factor ratios, which incorporate the
effects of residual stresses and stress ratio, to be determined based on Eq.4.3 1 or
Eq.4.36 depending on the crack depth and the depth of weld notch plastic zone. The
geometry correction factors a and Y are determined from the stress intensity factor
solutions developed in section 4.4.4.1 as

- 1'wa (Hp + - P) 1'


(4.39)
- Y (H P + 1 - p) ' C?:
4-71

where y1 is the average stress factor obtained form Eq.4.25 and is the bending stress
relaxation factor obtained from Eq.4.26. The coefficients H and Y are functions of
relative crack depth aft and crack aspect ratio a/c and are obtained from the stress
intensity factor solutions for a semi-elliptical crack given by Newman and Raju
(1981), and Y,, is the weld geometry correction factor obtained as a function of weld
geometry from BS PD6493 and Dijkstra et al (1990). The correction factor Y is
calculated corresponding to a constant crack depth of a=1.5mm. The crack
propagation equations in the two directions are coupled as the stress intensity factors
at the deepest point and at the surface tip are functions of the parameters alt and a/c.

Under variable amplitude loading, integration of Eq.4.37 in the thickness direction


from an initial defect size of a0 to a crack size a(t) after time t and average number
of stress cycles n(t) gives

a(s)
(4.40)
VL
C G(x) U' Y [nxlq]m12

where the LH.S. represents accumulated fatigue damage in time t and 1L(t) is called
a 'fatigue loading function' to be obtained as

n(t)
(4.41)
VL( t ) - E[ES m ] - E[n(t)]E[Sm]
i-I

Evaluation of this term depending on the long-term probability density function of


stress-ranges has been presented in detail in chapter 3.

Equation 4.40 needs to be integrated numerically using a computer program. Starting


from an initial crack depth a0 and initial semi-crack length c0, crack depth is
incremented in steps of, typically, 10-20% of its current value until the final crack
length is reached. Within each increment, the fatigue damage is calculated using an
8-point Gauss quadrature integration. For each increment of crack depth M, an
th
increment of crack length Lw and the crack aspect ratio a/c for the increment is
4-72

G(c) _____
Ac—
G(a) LUaYj
(4.42)
(a) — a11+Aa

When the crack depth reaches a value of 3mm the crack length is set to a value of
c=0.Old to account for coalescence of multiple-initiated embryo cracks. It is of
primary importance to evaluate the final crack depth a(t), which satisfies Eq.4.40, with
a close tolerance during gradient evaluations in an iterative reliability analysis
procedure. To achieve this, the residue of Eq.4.40, i.e (LH.S-R.H.S), is closely
monitored during crack growth analysis and incrementation is stopped when the
residue becomes positive. Using the lower and upper limits of crack depth in the last
increment as bounds on the root, a non-linear iterative analysis is carried out to
determine the value of the root a(t) which satisfies Eq.4.40 with a close tolerance
(residue <0.001). The corresponding semi-crack length c(t) can be calculated from
the knowledge of the crack aspect ratio.

When the fatigue life, Tf of a joint needs to be calculated, it is convenient to define


a fatigue resistance function of the joint as

a
dx (4.43)
VR(aC) -
2 CG(x) U Y[itx/q]m12

where a is the critical crack depth for fatigue failure of the joint taken here as the
plate thickness. It is possible to compute the fatigue resistance function almost
independently of the loading by incrementing the crack depth from an initial defect
size a0 to the critical crack size a. A fatigue loading rate I1L1(t= l year) can be
calculated from Eq.4.41 corresponding to an expected number of stress cycles in 1
year. Then the fatigue life of the joint (in years) can be calculated as

4-73

T1 - VR(aC)INfL1(t) (4.44)

4.4.4.4 Comparison with Experimental Results

The proposed crack propagation model is now compared with experimental results.
Detailed crack growth measurements have been reported for some of the tubular joints
tested at University College, London, see for example Dharmavasan and Dover (1982)
and Kam et al (1987). The geometiy and loading data for three of these joints used
for comparison are given in Table 4.1 below. The experimental data has been
obtained from Kam et al (1987).

Table 4.1: Geometry and loading data for the joints used for comparison with
experimental results

Stress
Joint Load T D t d
DoB Range
Type Type (mm) (mm) (mm) (mm)
Mpa
T OPB 16 457 12.7 324 0.889 150.0
Y IPB 16 457 12.7 324 0.825 150.0
X Axial 35 1228 19.0 750 0.800 140.0

* Assumed value

The predictions of equivalent Y-factors computed using the proposed model (Eq.4.39)
are compared with the compliance function curves obtained from experiments in
Fig.4.12. In general a good agreement can be seen for most part of the crack growth
for all the three joints. The predictions of the model are slightly conservative for the
initial phase of crack growth (aTF<0.3) because of the use of empirical weld geometry
correction factors developed from 2-D geometries and are known to be conservative
for tubular joints.
4-74

1.0

i.e
1.4

1.2

Vi

0.3

0.S

0.4

0.!

0
0.01 0.10 0.20 0.20 0.40 0.20 0.20 0.70 0.20 0.00 1.00
a/T

1.5

I.e

1.4

I.e

Vi

0.i

0.2

0.4

0.2

0
0.01 0.10 0.20 0.30 0.40 0.10 0.00 0.70 0.20 0.20 tOO
e/T

1.

I.

I.

1.

0.

0.

0.

0.

0
0.00 0.10 0.00 0.10 0.40 0.20 0.00 010 0.00 0.20 tOO
o/l

Fig.4.12: Comparison of Y-factor curves from the proposed model with experiments
4-75

0.S

0.S
alT
0.4

0.2

0.00 0.10 0.20 0.30 0.40 0.10 0.60 0.70 0.10 0.90 tOO
N/NF

Fig.4.13: Comparison of crack depth development from the


proposed model with experimental results.

0.2

0.6
a/c
0.4

0.2

0.00 0.10 0.20 0.30 0.40 0.10 0.80 0.70 0.20 0.90 tOO
a/T

Fig.4.14: Comparison of crack shape development from the


proposed model with experimental results
4-76

Similarly the predictions of crack growth in the thickness direction using the model
for the X-joint are compared with experimental results in Fig.4.13. Since all the
necessary parameters, such as weld geometry, initial crack size etc. are not available
it is not proper to compare absolute fatigue lives. The normalised crack growth curve
is seen to agree reasonably well with the experimental results for this joint.

The crack shape development predicted by the model is compared with experimental
results for the X-joint in Fig.4. 14 and the agreement is seen to be not so good.
However, it is to be noted that generally a large scatter is observed in reported results
on crack shape from experiments, see for example UEG (1985) and the model seems
to predict mean-value behaviour over a wider range of joint geometries. Crack shape
development when crack coalescence is not included is also shown in the figure. It
can be seen that the pattern of crack shape changes in this case is completely different
from that observed in experiments on tubular joints. It is therefore important to
include the effect of crack coalescence at least in an approximate way as proposed.

It is well recognised that fatigue crack growth in tubular joints is influenced by several
factors many of which are random in nature. From the seemingly good comparison
for few test results it should not be construed that the model will give accurate results
in every case. The main advantage of the proposed model is that it explicitly accounts
for most of the important factors which influence crack growth in tubular joints and
seems to predict the overall trend reasonably well. The proposed model is therefore
useful in a reliability analysis where the sensitivity of reliability of a joint can be
investigated with respect to a number of important parameters.
4-77

4.5 FRACTURE FAILURE OF TUBULAR JOINTS

Unlike fatigue, which is characterised by sub-critical crack growth under cumulative


cyclic loads, a fracture failure is characterised by unstable crack extension under a
single load application. Small weld defects in tubular joints may result in a crack of
significant size due to fatigue crack growth and if combined with a load of very high
magnitude may lead to fracture of the joint. The presence of tensile residual stresses
and the use of high tensile steels in deep waters or arctic conditions increase the risk
of fracture failure.

In the current practice, design against fracture is usually limited to specification of


appropriate material and welding requirements, see for example, API (1987) and D.En.
(1990). Material toughness is often determined through simple Charpy impact tests.
Explicit fracture assessment is used only to determine the "acceptability" of defects
of significant size found during fabrication or in-service inspections. However, in
view of the unavoidable nature of weld defects and their propagation by fatigue it is
desirable that every joint be assessed for fracture failure. In reliability assessments,
fracture should be considered as a potential failure mode because of the significant
sources of uncertainty involved in determining the fracture behaviour of tubular joints.
In the following, a number of theoretical and simplified design methods available for
crack instability assessment are reviewed and a simple methodology for fracture
assessment of tubular joints is proposed.

4.5.1 Review of Available Methods for Fracture Assessment:

Available methods for fracture assessment can be broadly grouped into two categories,
namely: linear elastic fracture mechanics methods (LEFM) and elasto-plastic fracture
mechanics methods (EPFM) depending on the extent to which a method accounts for
crack tip plasticity. The linear elastic fracture mechanics methods use either an energy
balance approach or an stress intensity factor approach of which the latter is now
widely used. Of a number of elasto-plastic fracture mechanics methods proposed in
4-78

recent years, the crack tip opening displacement (CTOD) method and the J-integral
method have found general acceptance. A number of simplified methods, based on
these two EPFM methods, have also been developed for routine design applications.
These methods are briefly discussed in the following under separate heads.

4.5.1.1 Energy Balance Approach:

Since early times several attempts were made to define a suitable parameter to
characterize the propensity of a crack to extend but the first successful attempt came
in 1920 when Griffith formulated the now well-known concept that an existing crack
will propagate if thereby the total energy of the system is lowered. He assumed that
there is a simple energy balance, consisting of a decrease in elastic strain energy
within a stressed body as the crack extends, counteracted by the energy needed to
create the new crack surfaces. Griffith's energy balance approach was originally
derived for brittle materials such as glass. Later, in the year 1948, Irwin modified this
approach to describe fracture phenomenon for more ductile materials. He pointed out
that the energy balance should be between: (i) the strain energy released during crack
extension and (ii) the surface energy plus the work done in plastic deformation. The
surface energy required to form new crack surfaces is generally insignificant in the
case of ductile materials. Irwin defined a parameter G, called the 'strain energy
release rate' or the 'crack driving force' which represents the elastic energy per unit
crack surface area that is available for infinitesimal crack extension. Irwin showed
that crack instability or fracture will occur when the strain energy release rate G
exceeds the resistance to crack extension R which for unit thickness can be stated as
4-79

G^R

G - L(FUa) (4.45)

R - •&(U.te+U.fp)

In the above,
U1 = change in elastic strain energy of the body
F = work done by external force
U = change in elastic surface energy due to new crack surfaces
U = work done in plastic deformation

It can be seen that for constant displacement or fixed grip condition F=constant and
G= -dUIda and for constant load condition FU a=Ua and thus G=+dUjda. The critical
value of the strain energy release rate G is considered to be a material property
which can be determined from experiments on simple specimens. Thus in a structural
component with a known crack, fracture occurs when the strain energy release rate G
due to applied loading exceeds the material value G.

Although Irwin's modification includes a plastic energy term, the energy balance
approach to fracture is still limited to an ideally sharp crack. Also, the energy balance
approach presents considerable problems for many practical situations, especially slow
stable crack growth, such as in fatigue. Thus, when there is some plasticity at the
crack tip and as long as the crack tip plastic zone is small when compared to the
dimensions of the crack, the stress intensity factor approach is now widely used which
is discussed next.

4.5.1.2 Stress Intensity Factor Approach:

In early 1950s, Irwin proposed the stress intensity factor approach which makes use
of the stress system in the vicinity of the crack tip to determine its propensity to
extend. Starting from Westergard's complex stress function and making use of

4-80
Cauchy-Riemann equations, Irwin showed that the elastic stress field on an element
in front of a crack tip can be expressed, in polar coordinates, in the form of

- K

r (4.46)

K YSm%I

where, K is the linear elastic stress intensity factor and is a function of the remote
stress Sm and the crack length a. The geometry correction factor Y is equal to unity
for an infinite plate with a centre crack of length 2a. The stress intensity factor thus
determines the magnitude of elastic stress field in the vicinity of the crack tip.

Irwin also demonstrated that if a crack is extended by an amount da, the work done
by the stress field ahead of the crack when moving through the displacements
corresponding to a crack length (a+da) is formally equivalent to the change in strain
energy G da. Thus the achievement of a critical stress intensity factor, Kc, is exactly
equivalent to the energy balance approach, which requires the achievement of a stored
elastic strain energy equal to Gc. The critical stress intensity factor K can be
determined from simple experiments on specimens. Thus if the stress intensity factor
K in a cracked component, due to any combination of applied stress and crack length,
exceeds the material fracture toughness K, it will result in unstable crack extension.

Using Eq.4.46, the distribution of elastic stress S, along the X-axis for 9=0 can be
obtained as shown in Fig.4.15. This shows that as r approaches zero, i.e. at the crack
tip, the stresses tend to infinity. However, most structural materials tend to deform
plastically above the yield stress and hence there will be a zone of plastically
deformed material surrounding the crack tip. Thus the elastic solution is strictly not
applicable in the immediate vicinity of the crack tip. Various attempts have been
made to estimate the extent of this plastic zone and to account for limited plasticity
within the theory of linear elastic fracture mechanics.
4-81

sy
Elastic Stress Distribution
K
f2itr

Sg

x
Crack Tip

Fig.4.15: Elastic stress distribution near a crack tip

Considering the stress distribution along the X-axis for 6=0 and substituting the yield
strength for S1 in Eq.4.46, a first estimate of the plastic zone size may be obtained as

I \
IlK (4.47)
r7 - - I-
2itIF,

However this does not take account of the re-distribution of stresses above the yield
value and hence the estimated plastic zone size will be smaller. Assuming a circular
plastic zone and ideal elastic-plastic material behaviour, Irwin derived the size of
plastic zone ahead of the crack tip as, see Ewalds and Wanhill (1985),

d, - 2r, - 0.318(K/F,) 2 (4.48)


4-82

Thus the plastic zone size according to Irwin is twice that obtained form Eq.4.47
without considering re-distribution. This is equivalent to assuming a notional crack
tip extending up to the centre of the circular plastic zone and the stress distribution
will be shifted by a distance r with respect to the elastic case.

S7
Elastic Stress Distribution

Stress Distribution

Fig.4.16: Effect of plasticity on crack tip elastic stress field

The elastic stress distribution S=K/f(2irr) takes over from F at a distance 2r ahead
of the actual crack tip as shown in Fig.4. 16. Later Dugdale derived an expression for
the plastic zone size using a Westergaard type stress function. Assuming that the
plastic deformation will be confined to a small strip ahead of the crack tip, the
Dugdale's 'strip yield model' gives, Ewalds and Wanhill (1985)

d - O.393(KF,) 2 (4.49)
4 83

This is slightly larger than the one given by Irwin.

Both the above approaches give a good estimate of the plastic zone size but assume
the plastic zone shape. Some attempts have also been made to obtain the plastic zone
shape but these methods assume the plastic zone size to be as given by Eq.4.47. The
approach is to express the elastic stress field equations in terms of principle stresses
and then use the von-Mises or Tresca yield criteria to give the boundaries where the
material starts to yield. These yield criteria do not account for the re-distribution of
stresses above yield and the consequent shift in the elastic stress distribution. More
exact evaluation of plastic zone size and shape using finite element methods shows
that the plastic zone size is smaller than that obtained by the von-Mises criterion for
the case of plane stress state while it is larger in the case of plane strain state.

The proposed expressions for plastic zone size show that the same value of K always
leads to the same value of plastic zone size and the stresses and strains, both inside
and outside the plastic zone, can thus be determined in terms of K. Hence the linear
elastic stress intensity factor can be used, even under crack tip plasticity, to correlate
crack growth and fracture behaviour. However, when the plastic zone size becomes
too large the predictions of LEFM become unsatisfactory and in such cases EPFM
methods have to be used. Turner (1984) defines four stress regimes to determine the
applicability of LEFM or EPFM methods. It is shown that LEFM methods can be
applied satisfactorily when the crack tip plastic zone is very small, such as in the case
of high tensile material under plane strain. When the plastic zone size is larger but
the plasticity has not spread to the lateral boundary of the component containing a
crack, EPFM methods can be used. However, in such situation of contained yielding,
LEFM with plasticity correction may also be satisfactorily used. But when the spread
of plasticity has reached the lateral boundary and is thus un-contained, the situation
is called gross yield and EPFM methods must be used. When the applied stress level
is beyond yield and plasticity develops over the entire section, the situation is called
4-84

general yield and this case cannot be handled by EPFM. In this case the component
may fail by plastic collapse and plastic limit analysis should be used.

4.5.1.3 Crack Tip Opening Displacement Approach:

For cases where significant plasticity develops near the crack tip, the plastic strain
controls the fracture process. A measure of plastic strain is the crack flank
displacement at or very close to the crack tip. Considering that the notional crack
extends beyond the actual crack tip, the crack tip opening displacement may be used
to quantify plastic strain. Thus it could be expected that at the onset of fracture the
crack tip opening displacement has a characteristic critical value for a particular
material and could be used as a fracture criterion.

The basis of the CTOD method lies in the Dugdale's strip yield model mentioned
earlier. Considering a notional crack of a+&i in which the length M is loaded with
uniform stress F, Dugdale argued that stress singularity, as predicted by the elastic
stress field Eq.4.45, cannot occur at the real crack tip as the stresses are limited to
yield stress. Thus to satisfy plasticity, the singularities from the local stress system
should cancel with the singularity from the applied remote stress system. Using this
approach Dugdale derived the expression for plastic zone size given by Eq.4.49.
Later, Burdekin and Stone (1966), using the Dugdale approach and expressing the
elastic stress field equation 4.45 in terms of displacements for the case of plane stress,
derived an expression for the crack tip opening displacement. The detailed derivation
is quite involved and can be seen in the said reference. The final result is

- 8Fa insec itS m (4.50)


2F

It is important to note here that the above expression for is derived for a centre
cracked infinite plate considering plane stress conditions. Furthermore, the material
is assumed to be elastic-ideal plastic. A similar result was also derived by Bilby,
Cotrell and Swindon (1963) for plane strain conditions and the model is popularly
known as the 'BCS-strip yield model'.
4-85
A major drawback of the CFOD method, which hindered its immediate practical
application, was that analytical expressions for similar to Eq.4.50 cannot be easily
derived for practical fmite geometry components. This in contrast with the stress
intensity factor and J-integral approaches for which K and J can be obtained from the
stress field ahead of the crack tip for a number of practical cases. Later Hayes and
Williams (1972) proposed a computational procedure using the Dugdale concept of
cancelling out stress singular terms to determine . Chell (1976) has extended the
BCS-Dugdale model to finite geometry plates subjected to arbitrary loading using
weight function techniques. Full 3-D elastic-plastic finite element analysis can be
used to compute 1 for any geometry and stress distribution, although at considerable
computational effort.

4.5.1.4 J-Integral Approach:

The J-integral approach is based on the energy balance approach discussed earlier.
Consider the energy balance given by

U - (U,,+U) - (Uo+UaF)+Uy (4.51)

in which
U = total energy content of a cracked body
U = potential energy of the body
U0 = elastic energy content of the loaded un-cracked body (a constant)
Ua = change in elastic strain energy due to the introduction of a crack
U.s, = change in elastic surface energy due to formation of new crack surfaces
F = work done by external forces.
This is valid for both linear elastic and non-linear elastic material behaviour. This can
also be extended to elastic-plastic material behaviour provided no unloading of the
cracked body occurs. This is because for actual plastic behaviour the plastic part of
the deformation is irreversible. Thus in the above equation, U contains all the energy
terms that may contribute to non-linear elastic behaviour while U is irreversible.
4-86
Following the approach in section 4.5.1.1 an instability condition for the crack is

d dU (4.52)
_(F-U ) ^ __L
da da

For the non-linear elastic behaviour and restricted elastic-plastic behaviour, the strain
energy release rate or the crack driving force can be denoted as J. Using Eq.4.5 1 and
noting that U0 is a constant, J can be defined as

I - i (F Ua) - - dU (4.53)
da --

Note that for linear elastic behaviour J=G by definition.

For a body of arbitrary shape containing a crack of length a and subjected to a


traction vector T over a small length ds along any arbitrary contour F, it can be
shown that, see for example Ewalds and Wanhill (1985),


I - fWdy-fTds (4.54)

where W is the strain energy density and is the displacement vector of the surface
of the body. In the above the change in potential energy is computed over a closed
contour and hence the method derives the name - J contour integral. Furthermore, it
can be shown that J remains constant for any arbitrary selection of the contour path
and is thus said to be path independent.

The path independency of the J integral has an important significance because now
J can be computed along a contour remote from the crack tip, chosen in such a way
as to contain only elastic stresses and displacements. Thus, interestingly, crack
instability which is controlled by plastic stress-strain field close to the crack tip can
be predicted by purely elastic computations. This is what makes the J integral
approach so powerful. The path independency of the J integral has been confirmed
by a number of numerical studies by choosing contours both within and outside the
plastic zone. If the parameter J is assumed to represent crack tip conditions correctly,
4-87

then for every material there should be a critical value of J, namely J, at which
fracture occurs.

The J integral can also be given an alternative interpretation as a parameter


characterizing the stress field singularity at the crack tip as shown by Hutchinson, Rice
and Rosengren (c.f. Turner 1984). They studied the stresses and strains in a
non-linear elastic material and derived the power of singularity in stress, strain and
strain energy density. These solutions are popularly referred to as HRR solutions.
Later Mclintock (c.f. Turner 1984) expressed the stress field equations near the crack
tip in terms of J.

It is to be noted that J, as discussed here, was initially derived for 2-D stress-strain
field for non-linear elastic material. It can be applied to elastic-plastic materials under
monotonic loading. Thus, strictly the J concept is applicable only up to the point of
initiation of crack extension. However, because of its interpretation as a stress field
characterising parameter it can also be used to describe ductile tearing and stable crack
growth. The original energy balance formulation did not include residual and thermal
stresses or the crack face loading between the crack tip and the points at which the
contour of integration cuts the crack flanks. However, a number of later developments
have made the J integral concept applicable for general 3-D stress states including the
effects of secondary loadings.

The four theoretical concepts discussed so far are essentially equivalent. For linear
elastic conditions it is possible to show that the four fracture mechanics parameters
G, K, 8 and I are interrelated as, for plane stress


G - I - ö,F - K2/E (4.55)

For the elastic-plastic case J and ö are related in a general form given by

J - M,F, (4.56)
4-88
where the value of M is found to lie between 1 and 3 with lower values for membrane
loading and higher values for bending loading. The value of M is influenced by
factors such as plastic constraint, work-hardening, extent of plasticity in relation to the
geometry of the body and the way is measured.

Although the CTOD and J integral methods provide a very good description of the
fracture process, they are usually very difficult to apply for practical geontries and
stress states for which simple analytical solutions for or J are not available at
present. These have to be computed using finite element analysis which can be quite
involved. A number of simplified methods have therefore been developed for design
applications which make use of either the CFOD or the J integral concepts. Only two
of these methods which have found wider application are reviewed in the following.

4.5.1.5 BS PD 6493 Approach:

This is one of the most widely used methods of fracture assessment in the offshore
industry. This was proposed to provide guidance on derivation of acceptance levels
for defects in fusion welded joints. The first edition of the document, BSI (1980), was
based on the so called COD design curve derived from few wide plate tests. The
basis of the method lies in the analytical design curve derived by Burdekin and Stone
(1966) resulting in an expression for \ as given by Eq.4.50 for centre cracked infinite
plate under plane stress. In order to develop a general design procedure, they defined
a dimensionless CTOD parameter as

S
- ____ ( 4.57)
2itFa 2tca

where is the elastic strain at the yield point. In the original analytical
method, a number of curves, for various gauge distances (a/y), expressing (I) as a
function of the relative strain EIc were derived using the Dugdale strip yield model-
to compute elasto-plastic displacements and strains at a number of points in the
vicinity of the crack. A number of experiments conducted in the late sixties, however,
4-89
indicated that cb is almost independent of (aly ) . Therefore, the analytical curves were
replaced by a single empirical curve defined as an upper bound to the test data. This
is the curve adopted in BSI (1980). With a further modification from Dawes (1980),
who argued that for small cracks (a/W)^0.1 and applied stresses below yield,
(2/Cy)(SmfFy), the final design curve may be expressed as

m 6E
for Sm /Fy <0.5
- 2itF,a
- (4.58)
81E
cb - .-0.25 - for Sm/Fy ^ 0.5
F 27tFa

The maximum permissible crack size, am, can be obtained directly from the above
equations by substituting the critical CTOD value, , as

&:CPUEFY
a
mi - for SJ /F <0.5
2itS
(4.59)

a - _______ for 0.5 <S1/F < 2


2it(S1-0.25F)

where S 1 is a sum of all stress components including applied nominal stresses,


additional stresses due to stress concentrations and secondary stresses such as welding
residual stresses.

The COD design curve was only intended to provide "acceptable" defect sizes and not
"critical" crack sizes for a known service loading and was thought to be highly
conservative. Comparing 8 from Eq.4.59 with that obtained from Dugdale analysis
Eq.4.50 it can be seen that the COD design curve has an inherent bias of 2 for
(S1/F)<0.5. However, recent studies, see for example Rhee (1985), have shown that
the bias varies for (S 1/F)>0.5 and the design curve even becomes un-conservative at
high values of applied stress. This variable bias inherent in the model makes it
unsuitable for use as a mechanical model in a reliability analysis. Another limitation
of the method is that the design curve was derived from wide plate tests under
membrane loading and it does not explicitly account for stress variation through the
4-90

thickness. Anderson et al (1988) have suggested modifications to the design curve to


remove some of the limitations and suggest that the design curve be used for applied
stress S<O.85Fy . The realisation of above limitations of the design curve has led to
the recent revision to PD 6493 guidance and a draft document, BSI (1989) is currently
under review by the industry.

The proposed revised document offers the user 3 levels of assessment methods, with
increasing complexity and better accuracy. The Level 1 is very similar to the original
COD design curve and is only used for preliminary screening of welds. The Level 2
method uses a so called 'universal failure assessment diagram approach' which is
discussed in detail here. The Level 3 method is used when more detailed material
stress-strain data for the work-hardening part is available or ductile tearing prior to
fracture is to be considered.

1.2

1.0

0.8

0.6

04

0.2

0
0 0.2 04 06 0.6 1.0 1.2
Sr

Fig.4.17: BS PD 6493 Failure Assessment Diagram, BSI (1989)

In Level 2 method, a universal failure assessment diagram, as shown in Fig.4.17, is


assumed to be applicable for all material types and cracked geometries. The failure
assessment curve is defined by
4-91

(4.60)
- 5r[4h'i
sec(!.SrJI

This curve is basically derived from the Dugdale's strip yield model, Eq.4.50. The
parameter S7 is defined as


Sr - (4.61)
flo

where S is the stress in the remaining cross section of a cracked component and Ffl0
is the material flow stress, taken generally as the average of yield stress and ultimate
tensile stress. The parameter S7 thus determines the closeness of a cracked body to
plastic collapse. Then for the cracked body a fracture parameter 6r due to applied
loading is determined as

I \112
app1 + (4.62)
PC

where 6 is the material toughness in terms of critical CTOD and p is a correction


factor to account for the interaction of primary and secondary stresses and is
determined from empirical equations given in the reference. The crack tip opening
displacement due to applied loading is determined from


2
app - (4.63)

where K is the stress intensity factor due to combined primary and secondary stresses,
a classification based on whether a particular source of loading contributes to plastic
collapse or not. It is interesting to note that applied crack tip opening displacement
is obtained from linear elastic stress intensity factor while the elasto-plastic
considerations are treated using the failure assessment curve. The revised document
also allows the entire assessment to be made in terms of K if the material data is
available in terms of fracture toughness K instead of 8,. In this approach it is
implicitly assumed that, although the failure assessment curve is based on a centre
4-92
cracked infinite plate subjected to uniform tensile loading, the effects of specific
geometry, stress distribution are incorporated through the parameters and Sr

4.5.1.6 CEGB R6 Approach:

This method has been developed by the U.K. Central Electricity Generating Board
(CEGB). The original method proposed in 1980 (R6-Rev2) was essentially the same
as the Level 2 method of latest PD 6493 document except that it used only the stress
intensity factor format. However, this method has been replaced in the revision 3
issued in 1986, see Mime et al (1985), which is based on a J integral analysis. The
Rev-3 of the R6 approach offers three categories of assessment methods depending
on the type of fracture toughness data available and the extent to which ductile tearing
resistance is to be considered. The Category 1 method evaluates fracture initiation
using the material data on initiation fracture toughness. While in the Category 2
method fracture assessment is performed for two values of fracture toughness, namely
at K, the fracture toughness conesponding to 0.2mm of tearing, and at K0, the
maximum value of toughness available from a full tearing resistance curve. The
Category 3 analysis uses the complete tearing resistance curve and the analysis is
performed at various points on the curve with the corresponding values of crack
extension (ia) and fracture toughness K(i.a).

For each category of analysis, three options of failure assessment diagrams are
available. The option 3 failure assessment curve is developed from a detailed J
integral analysis of the cracked component taking due account of the component
geometry, stress distribution and material strain hardening behaviour. This option
gives the best accuracy but is considerably more difficult to use. For practical
applications an Option 2 failure assessment diagram is proposed in which the
geometry dependence is removed using conservative assumptions. However, the
Option 2 method uses a detailed material stress-strain behaviour and a specific failure
assessment curve is developed for each material. This method therefore gives good
results for materials with high strain hardening behaviour. The Option 1 is a universal
4-93
failure assessment diagram for all material types for which the failure assessment
curve has been developed to give a lower bound to assessment curves of different
material types. The use of this option is discussed in detail in section 4.5.2.

In addition to the above two methods, the EnJ method given by Turner (1984) and the
EPRI method, see Turner (1984), have gained wider recognition. The EnJ method
gives an empirical design curve relating J with the uncracked body strain. The curve
is based on computed results for various crack and structural geometries. The EPRI
method is based on a detailed J analysis for each cracked component. Standard fully
plastic solutions for different crack and component geometries are being developed
which can be directly compared with the material J data. Alternatively, the J
solutions can also be plotted in the form of a failure assessment diagram. The method
results in a family of failure assessment curves for different crack configurations and
a given stress-strain material curve. The method is very rigorous but standard
solutions for all practical cases of interest may not be available in the near future.

4.5.2 Proposed Methodology for Fracture Assessment of Tubular Joints:

Although the CTOD and J integral methods are the most accurate methods for
elastic-plastic fracture assessment, their application to tubular joints will need detailed
3-D finite element computations which may not be possible in routine design
applications. Of the simplified design methods discussed before, the CEGB R6
method is considered to be the most appropriate for use in a reliability assessment of
tubular joints. The recent revision of PD 6493 brings it very close to the R6 method
although there are some differences in the failure assessment curves employed by the
two methods. The salient features of the R6 method and its application to tubular
joint problems is discussed in the following.

4-94

4.5.2.1 Failure Assessment Diagram:

A universal failure assessment diagram is chosen here for reasons of simplicity. The
use of a material specific failure assessment curve, although in principle suitable for
use in a reliability analysis, is not considered as enough data may not be available to
develop a probabilistic description of the entire material stress-strain curve. It may,
however, be possible to assume a material strain-hardening law, such as a
Ramberg-Osgood law, the parameters of which may be considered as random. But
this will dilute the marginal accuracy that could be gained from the use of a material
specific failure assessment diagram. On the other hand, a more rigorous J based
failure assessment diagram for each component and stress distribution is considered

K,1 (1014L,2)[O3+O7exp(_O65L6)}

0.8 .
K,.
0.6 .- .. .........................................

0.4 .

0.2 . S ..

- i •1 I j
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4
Lr

Fig.4.18: CEGB R6 Failure Assessment Diagram, Mime et al (1985)

to be unsuitable for direct use in a reliability analysis for computational reasons. The
universal failure assessment curve as given by R6 Rev-3 is, see Milne et al (1985),

K 1 - [1 - 0.l4L] [0.3 + 0.7 exp(-0.65L)] (44)


4-95

where L7 is a non-dimensional collapse parameter and K 7 is a non-dimensional fracture


parameter. denotes the value of K 7 at failure. The failure assessment diagram is
shown in Fig.4.18. The failure assessment curve is a continuous line varying from
K11=l at L=O to K11=O for a very large value of L1.. However, for high values of L7
the cracked component will be in a situation of general yielding and may fail by
plastic collapse instead of by fracture. For situations of general yielding EPFM
analysis is not satisfactory and hence a cut-off is imposed on the failure assessment
curve at L=L , beyond which K.f. is taken as zero. The value of is generally
chosen to correspond with plastic collapse of the cracked joint. Thus the failure
assessment curve describes the fracture behaviour of a component ranging from linear
elastic brittle fracture at one extreme, ie at L 1 =O to plastic collapse at the other
extreme, ie at L7=L,1 . It is important to note here that this failure assessment curve
was a derived as a lower bound to a number of material specific failure assessment
curves and is thus expected to be conservative. Furthermore, the curve is essentially
derived from J integral calculations unlike the PD 6493 Level 2 assessment curve,
shown in Fig.4.17, which is based on a strip yield crack tip opening displacement
model. One other limitation of the failure assessment diagram chosen here is that it
has been developed for essentially 2-D geometries, such as flat plates, and their use
for 3-D geometries such as tubular joints has not yet been validated. However, the
geometry dependence of the curve is expected to be small and if appropriate account
of the joint geometry and stress distribution is taken in calculating the parameters L7
and K7, as in the following, the use of the curve may be acceptable.

4.5.2.2 Plastic Collapse Parameter Lr:

All the loads acting on a cracked joint are first classified as those contributing to
plastic collapse (called primary loads or stresses) and those that do not contribute to
plastic collapse (called secondary loads or stresses). Only those loads contributing to
plastic collapse are included in the evaluation of L7. Thus, welding residual stresses
which are self-equilibrating are considered as secondary stresses and not included in
4-96

the evaluation of L1. However, they are included in the evaluation of K 1 as they
contribute to the total stress causing fracture. The parameter L is defined as

load contributing to plastic collapse (4.65)


L-
plastic yield load of the flawed structure

Thus the parameter Lr determines the closeness of a cracked component to yielding.

In order to calculate L1 it is necessary to calculate the plastic yield load of the joint,
which is defined in general terms as the load required to cause yielding of the
remaining cross section of a cracked component considering elastic-ideal plastic
material behaviour. For part-through cracks, the R6 document recommends the use
of "loca1 yield load, ie the load required to cause plasticity to spread across the
remaining ligament below the crack. However, the use of local limit load is likely to
be highly conservative for tubular joints. Because of the high stress concentration the
hot-spot stress at the crack location will reach a yield value much before the final
failure of the joint by collapse. Moreover, because of the high restraint available in
tubular joints the ligament yielding will be contained and the joint will be able to
support much higher loads. For these reasons it is suggested that plastic yield load
should be determined from the collapse load of the joint. Since joint failure by plastic
collapse is based on combined loads acting through the brace, based on Eq.4.7, a
parameter Sr is initially determined as

M12 +
S - . 1+ (4.66)
P1
l)
MI
M,/JPB
M
' OPB

where P, M etc are the collapse loads for the joint, under individual loading modes,
reduced to account for chord loads and presence of the crack as discussed in sections
4.2.5 and 4.2.6. These loads, however, correspond to collapse of the joint and include
strain hardening. In order to use plastic yield loads, the parameter L1 is obtained as

F+F
' (4.67)
L-
2F,
4-97
Here it is assumed that plastic collapse of a joint occurs conesponding to material
flow stress which is taken as the average of the material yield stress and ultimate
tensile stress. For the same reasons, the maximum value of the collapse parameter
L, which defines the cut-off of the failure assessment curve, is taken as

F+F
L-
rm
(4.68)
gg

The use of the "global" limit load as above is expected to give a better reflection of
the closeness of a joint to general yielding condition. However, it may not give a
better quantification of crack tip plasticity condition which was one of the reasons
why a limit load parameter was introduced into the J-integral analysis used in
developing the failure assessment diagram, see Ainsworth (1984). The use of either
a local or global yield load for tubular joints needs to be validated using experiments
or detailed J integral analyses.

4.5.2.3 Fracture Parameter K:

The fracture parameter K is evaluated for the combined primary and secondary
stresses as

K-

K" - K"
(4.69)
"mat

r, S
S At
it,. -

mat

in which K is the linear elastic stress intensity factor and K. is the material fracture
toughness. The superscripts 'p' and 's' denote primary and secondary stresses,
respectively. P is a correction factor to account for the plastic interaction between
primary and secondary stresses. This correction is usually small and can be obtained
from the equations given in Mime et al (1985). A simplified model for evaluation of
stress intensity factors K for tubular joint cracks has already been presented in
4-98
section 4.4.4.1. The crack size used here may be the initial defect size or a postulated
crack after fatigue propagation during service.

It is interesting to note that the R6 method, although an elastic-plastic fracture


assessment method requires only the knowledge of elastic stresses and linear elastic
stress intensity factor while all the elastic-plastic considerations are included in the
failure assessment curve. This is one of the most important advantages of the failure
assessment diagram approach and offers the user a considerable degree of simplicity
when compared to methods which require explicit calculation of either the crack tip
opening displacement or the J integral for a given cracked geometry.

4.5.2.4 MateriaL Fracture Toughness K:

The representative values of fracture toughness to be used in the assessment are K1c,
Kc or K02 defined as follows:
K1 the linear elastic, plane strain fracture toughness at the onset of brittle fracture,
meeting fully the requirements of BS5447 (1977) or ASTM E399 (1979).
Kc : the fracture toughness at the onset of brittle fracture where the validity
requirements of BS5447 or ASTM E399 are not fully met.
K02: the fracture toughness after 0.2mm of crack extension in a J controlled crack
growth test given by CEGB R6; Appendix-i, see Mime et al (1985). This
gives a good estimate of the initiation of ductile tearing.

It is suggested that full benefit of ductile tearing resistance should not be considered
for assessment of tubular joints of offshore structures since the fracture process is
usually accompanied by fatigue crack growth. In an extreme storm situation there will
be a number of stress cycles of similar magnitudes and fracture onceinitiated under
one of the peak cycles will be coupled with very high fatigue crack growth per cycle
and may well lead to the complete failure of the joint. This has in fact been observed
in some tests on tubular joints, see Gibstein and Moe (1986), where the test specimens
failed completely after repeated application of the load corresponding to initiation.
4-99
For this reason, fracture once initiated is considered to cause complete failure of the
joint and no allowance is made for crack arrest which may occur under single load
application.

The choice as to which of the three toughness parameters to be used in an assessment


depends on a number of factors. Materials can vary in their fracture behaviour from
fully brittle to fully ductile. The type of behaviour depends on variables such as
material microstructure, prevailing stress state at the crack tip, operating temperature
etc. Ductile behaviour is expected in low tensile materials and thin sections where
plane stress conditions are expected and brittle behaviour is associated with high
tensile materials in thick sections where plane strain conditions are expected. Fracture
behaviour is strongly influenced by temperature with pure brittle fracture, sometimes
called cleavage fracture, occurring at veiy low temperature. At high operating
temperatures the fracture behaviour is predominantly ductile and the transition
temperature regime showing a mixed brittle/ductile behaviour with toughness values
increasing sharply with temperature. For each material the transition temperature
should be established by careful experiments. Material microstructure also has a
strong influence on fracture behaviour with materials of higher grain size exhibiting
inter-granular fractures with little plasticity. The specimen used for fracture testing
should therefore be taken from the appropriate material, ie weld metal, HAZ or parent
plate, depending on where the crack tip is located.

Depending on the service temperature, stress state and material type of the actual
component, the type of fracture failure mode, ie brittle or ductile, should be
anticipated and accordingly material properties K1 or K should be used if brittle
fracture is expected and K 2 to be used when a ductile mode is assured. Once a
decision is made on the relevant toughness parameter to be used, appropriate specimen
thickness and test procedures given by the relevant testing standards should be
followed to obtain the material data. Whenever possible specimens of the same
thickness as the component under investigation should be used and the test procedure
should take account of the orientation of the flaw with respect to the loading in the
actual component. When weld metal is being tested the specimens should be obtained
4-100

from the samples used for weld qualification of the particular weld. When HAZ
material is being tested, consideration should be given to the changing material
microstructure with depth and the depth at which the crack tip is located in the actual
component A record of the load-displacement during testing should be kept and
should be used in interpreting the test results based on guidelines given in CEGB R6
document or elsewhere.

Although, the material properties required in the analysis are in K format and should
preferably be obtained by K testing, it is possible to convert values obtained by other
test methods to K format. When J testing is used, for example in determining J 02, it
can be converted to K format using, see Milne et al (1985),

EJ
O.2 (4.70)
1-v2
where 1 is the poisson's ratio. When CTOD testing is used according to BS 5762
(1979), and if the test result corresponds to brittle failure 6, or initiation of ductile
tearing ö, these values may be converted to equivalent K values using, see Ewalds
and Wanhill (1985),
For plane stress conditions:


K - EF& (4.71)

For plane strain conditions:

(4.72)
K- ______

where ? is a 'plasticity constraint factor' varying between 1.5-3.0. However, if the


CTOD results are defmed with respect to the maximum load after significant ductile
tearing, 6m' or corresponding to crack instability after ductile extension, ö, these
values cannot be converted to equivalent K. values and in such a case it is better to
use the P1)6493 Level 2 failure assessment diagram, using the 6 route, as this diagram
is essentially derived from a CTOD model. For some of the older platforms toughness
values may be available only in the form of Charpy impact data and in such cases
they can be converted to equivalent K 1 c values. A number of empirical correlanons
4.101

betWeefl Charpy V and K1 are available, see BSI (1989) for appropriate references,
and they should be used with care taking due account of the uncertainties involved.

Following the approach described above, a cracked component is considered to be safe


if the coordinate point (KrLr) due to applied loading falls within the failure
assessment diagram. If KII^KI, it implies that the component fails by fracture while
I implies plastic collapse. In a reliability assessment these two conditions
should be treated as separate, but correlated, limit-states.

4.5.2.5 Comparison with Experimental Results:

The CEGB R6 failure assessment diagram has been validated against results from
numerous standard specimens and few pressure vessel structures, see Ainsworth et al
(1986). However, validation of this model for tubular joints has not been established
so far. Brittle fracture tests on tubular joints have been reported by Gibstein and Moe
(1986) and Machida et al (1987). The fracture stress obtained from these tests along
with other relevant data has been used to calculate Lr and KT parameters based on the
proposed model for tubular joints. The important input data used is given in Table 4.2
where points 1-6 correspond to Gibstein tests and the remaining two correspond to
Machida tests. The results are also plotted on the failure assessment diagram as
shown in Fig.4.19. In evaluating the test results from Gibstein and Moe (1986)
appropriate values had to be assumed for yield stress and residual stresses as the
authors have not reported these values. The material properties reported in terms of
initiation CTOD values have been converted to K values using Eq.4.71. Plane
stress conditions have been assumed in view of the relatively small crack depths. On
the other hand, in interpreting test results from Machida et a! (1987) plane strain
conditions have been assumed in view of deeper cracks and Eq.4.72 has been used
with a constraint factor of 1.5 for obtaining equivalent K. values. In view of these
uncertainties in input the results should be viewed with some scepticism.
4-102

Table 4.2: Comparison of the proposed fracture assessment model


with experimental results.

SI. Hot-spot Residual


No Stress Stress a1 Lr K1.f
N/mm2 N/mm2 mm mm N-
mm312
1 518.40 300.00 4.0 100 2962 0.372 1.03 0.98
2 717.80 300.00 3.0 100 2962 0.511 1.17 0.95
3 691.20 300.00 5.4 47 2796 0.491 1.33 0.96
4 581.55 300.00 4.0 14 3042 0.409 0.93 0.97
5 664.62 300.00 3.0 100 3873 0.473 0.87 0.96
6 598.16 300.00 3.8 100 3873 0.428 0.86 0.97
7 253.00 200.00 10.0 80 1989 0.135 1.06 0.99
8 522.20 50.00 10.0 80 2089 0.282 0.99 0.98

C ______________________
1.2-.
C
o Gibstein and Moe (1986)
• C: .....(___ _
St()
1

0 .4 ..- ....:. ...............................................

0.2 .0 - . L-.-....-- _____________________________________________


0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4

Fig.4.19: Comparison of the proposed fracture assessment model


with experimental results

From these results it can be seen that the proposed model gives a reasonably good
estimate of the mean fracture load. Some of the points lie below the failure
assessment curve and it is therefore necessary that adequate safety factors should be
considered, for important variables such as fracture toughness, applied suess, crack
size etc, for using this model in a deterministic design, see BSI (1989, Appendix-i).
Chapter 5
RELIABILITY ANALYSIS OF TUBULAR JOINTS

5.1 GENERAL

Chapters 3 and 4 presented mechanical models required to obtain load induced


response of jacket structures and to estimate the strength of individual tubular joints
in the structure. These models essentially constitute the limit-state function for tubular
joints of jacket structures. Thus, given the input values for all design quantities, this
limit-state function gives a discretised assessment of the state of each tubular joint in
a structure as failed or safe. However, in order to account for the uncertainties in
these design quantities they should be treated as random variables, prescribed by
appropriate probability distributions, and the probability of failure should be calculated
for individual tubular joints.

In this chapter, the reliability analysis of tubular joints is considered with respect to
joint failure by fatigue, plastic collapse or fracture. Various sources of uncertainty
involved in the estimation of structural response in terms of stresses induced at tubular
joints and in the strength prediction of joints are identified and appropriate methods
for the treatment of these uncertainties in a reliability analysis are discussed. Using
the loading and strength models developed in previous chapters, a methodology for
reliability assessment is presented for the limit-states of fatigue failure, fatigue-plastic
collapse, and fatigue-fracture.

A detailed system reliability analysis of a jacket structure in 60m water depth has been
carried out using the RASOS software. The geometry of the structure and
environmental conditions used are given in Appendix-D of the thesis. Results of the
reliability analysis for the complete structure are presented in chapter 6 in the context
of a system analysis. In this chapter only a few selected joints are studied in greater
detail and the results are presented in the form of a number of examples. Based on
these example reliability analyses, the most important random variables are identified
and the sensitivity of the probability of joint failure to these variables is determined.
5-2
Failure probabilities are computed using the first order reliability method (FORM)
discussed in chapter 2. From these results a number of important factors that need to
be considered in a system reliability analysis under fatigue conditions are identified.

As discussed briefly in chapter 1, uncertainties in the failure prediction of tubular


joints can arise from a large number of sources all of which may not be amenable for
systematic analysis and estimation. In a reliability based assessment, at least the
important sources of uncertainty should be recognised and accounted for through a set
of basic random variables. The uncertainties in the assessment of tubular joints may
be broadly categorised as involved in the: (i) estimation of extreme environmental
loading, (ii) estimation of fatigue loading, (iii) strength prediction models and (iv)
material properties. Treatment of these uncertainties are discussed under sections
5.2-5.5 while reliability analysis for different limit-states are presented under sections
5.5-5.8 in the following. The main conclusions from these reliability analyses are
summarised in section 5.9.

5.2 UNCERTAINTIES IN EXTREME ENVIRONMENTAL LOADING

As discussed in chapter 3, computation of internal forces in the primary members of


a jacket structure involves a series of steps which can be broadly grouped into: (i)
environmental description, (ii) calculation of hydrodynamic loading and (iii) global
stiffness/modal analysis to determine nodal displacements arid forces. Uncertainties
can arise in each of these steps and ideally each source of uncertainty should be
modelled through a separate basic variable. To some extent, the modelling of
uncertainties depends on the method of response analysis employed as this also
dictates the type of environmental description. Uncertainty modelling in this section is
discussed in the context of discrete wave deterministic-static analysis as this is the
commonly used extreme response analysis method for shallow-water jacket structures.
A spectral method is commonly used for fatigue analysis and uncertainty modelling
in this case is discussed in the next section on fatigue loading. The uncertainty
analysis under extreme wave event is discussed under the following three heads.
5-3
5.2.1 Environmental Description:

Some of the important environmental parameters considered in the response analysis


of jacket structures are: significant wave height and mean zero-crossing period for the
storm sea-state, wave height and wave period of the highest wave, water depth, wind
velocity and current velocity. Ideally a joint description of wave, wind and current
conditions should be developed and the extreme load event should be considered as
the set of environmental parameters which cause the highest loading on the structure.
However, as discussed in chapter 3, it is common to assume that the highest structural
response occurs when a wave with the life-time extreme wave height passes through
the structure. Analogously, in a frequency-domain analysis, the highest response is
assumed to occur in a storm with the life-time extreme significant wave height.
Furthermore, it is often assumed that the life-time extreme wave height occurs in the
life-time extreme storm. The wind velocity and current velocity can be modelled
conditional on the extreme significant wave height. Uncertainty modelling for each
of these important environmental parameters is given below.

Extreme Si gnificant Wave Hei ght H


The total uncertainty in extreme significant wave height comprises: inherent
randomness, statistical uncertainty and model uncertainty. Each of these sources
should be analyzed separately and, if necessary, combined together to obtain the total
variability. Uncertainty model for the inherent variability in wave conditions is often
developed from wave hindcasting. A Gumbel distribution is commonly used to model
the inherent variability in extreme significant wave height. For example, Olufsen and
Bea (1990) have analyzed hindcast data for the period 1955-86 in the northern North
Sea and obtained values of 7.55m for the mode (u) and 1.2055 for the scale
parameter (a).

The statistical uncertainty, arising due to limited data, in the estimated distribution
parameters (assuming the Gumbel model is valid) can be obtained by analyzing data
sets from different locations in the same field. This uncertainty can be modelled by
treating the distribution parameters u and a for the extreme significant wave height
5-4

themselves as random variables. The Olufsen and Bea (1990) data set for the northern
North Sea gives a standard deviation of 0.12m for u and 0.0723 for a with the
corresponding mean values given as above. Analysis of the data also show that the
parameter u can be modelled as a normal variable and a as a lognormal variable with
correlation of 0.9 for the two variables. Note that the above model parameters are
applicable only for the data-set considered and should not be used for other locations.

Significant uncertainties can be involved in the model used for wave hindcasting and
these can be estimated by comparing the hindcast data with measured data at the site.
For the hindcast model used for the analysis of northern North Sea data Olufsen and
Bea (1990) have found that a model uncertainty variable B11, giving the ratio of
measured to hindcasted significant wave height, can be modelled as a lognormal
variable with a mean of 0.95 and standard deviation of 0.15. This variable was found
to be relatively independent of the inherent variability in significant wave height.

The use of a Gumbel model for describing the variability in extreme significant wave
height can also introduce some uncertainty which is often difficult to quantify. The
use of a 2-parametered distribution to fit sample moments gives a reasonably good fit
in the body of the distribution but their tail behaviour beyond observed values is
largely subjective. Winterstein and Haver (1990) have proposed a generalised Gumbel
model in which the higher moments, such as skewness and kurtosis, from the sample
are used to adjust the coefficients in the model to get a good fit in the tail region.
Thus the model uncertainty is greatly reduced while the statistical uncertainty in the
estimation of higher moments is slightly increased which can be readily quantified.
This method should therefore be preferred when the available sample data is large, so
that higher moments can be estimated with close confidence.

Mean Zero-crsn Period Tj


The mean zero-crossing penod for the extreme sea-state can be defined conditionally
on the extreme significant wave height. A lognormal distribution has been found to
give a good fit to the data, see for example Haver and Winterstein (1990). In this
5-5
case the conditional mean is assumed to be linearly related to the significant wave
height as pTZ=ATZ+BTZ H. On the other hand the variance is assumed to be
independent of significant wave height. As an alternative to this approach, as
mentioned in chapter 3, the characteristic wave steepness can be modelled as a
lognormal variable independent of significant wave height.

Extreme Wave Hei2ht H:


There are two alternatives available to model the variability in extreme wave height.
In the first approach extreme wave height can be defined conditional on the extreme
significant wave height assuming that extreme wave occurs during the extreme storm
sea-state. The limitations of this approach have already been discussed in chapter 3.
In this case the distribution function of extreme wave height conditional on the
extreme significant wave height can be derived. The associated wave period T of the
extreme wave is usually assumed to be a deterministic function of mean zero-crossing
period (T=1.2T). The statistical and model uncertainties are therefore introduced
through 113 as previously discussed.

The second approach consists of developing hindcast data directly for extreme wave
heights in which case it is not necessary to include extreme significant wave height
and mean zero-crossing period as basic variables in a reliability analysis. The
statistical uncertainties and model uncertainties associated with the hindcast model
should also be determined based on data from wave heights. The associated wave
period is defined conditional on the extreme wave height using some deterministic
relationship, for example using constant wave steepness. Olufsen and Bea (1990) give
uncertainty statistics for extreme wave heights for the north-west Gulf of Mexico wave
climate. The same approach has been adopted by Carr and Birkinshaw (1988,
01'! 537) in a study conducted for the Department of Energy.

Water Depth d:
The uncertainty in the actual water depth at the offshore location arises due to three
main sources. The first is in the uncertainty about the measured water depth due to
uncertainties in echo sounder accuracy, ship motion, positional accuracy etc. This can
5-6

be subjectively modelled by treating measured water depth as a normal variable with


a standard deviation of up to im, Can and Birkinshaw (1988, OTI 537).

The second source of uncertainty arises in the correction applied to measured water
depth with reference to latitude. The level of uncertainty depends on the survey
procedure and whether adjustments are made for non-tidal variations at the site using
a tide gauge. When no such adjustments are made and the latitude correction is
applied based on tide-tables for the nearest standard port, which is often the case, Can
and Birkinshaw estimate that the maximum error could be about plus or minus 1.2m.
Thus the latitude correction to measured water depth can be modelled as a normally
distributed variable with zero mean and standard deviation of 0.5m.

The third factor which affects the water depth at a location is the local surge level
which has to be estimated from actual measurements. Haver and Winterstein (1990)
suggest that the surge correction to water depth should be modelled conditional on the
extreme significant wave height (or extreme wave height) as a normal variable with
a constant standard deviation and a mean linearly varying with the actual significant
wave height (or extreme wave height).

In a reliability analysis it is possible to treat each of the three sources of uncertainty


as independent basic random variables in which case the contribution of each source
to total uncertainty can be studied. However, initial studies show that the total
uncertainty due to all the three sources is typically about 5-8% and for computational
reasons it is preferable to combine the three sources of uncertainty. In this case a
random variable M, which is a correction to the measured water depth can be
modelled, conditional on the extreme significant wave height (or extreme wave
height), as a normally distributed variable with a constant standard deviation of
typically 2.Om and the mean depending on the local surge level.

Wind S peed V:
The total uncertainty in wind speed, similar to wave height, is expected to comprise
inherent variability, statistical uncertainty and model uncertainty. Although each of
5-7
these three sources can be analyzed and modelled as three separate basic variables, in
view of the secondary importance of wind loading compared to wave loading on
jackets it is preferable to model wind speed through a single basic variable. When
adequate data is available it is preferable to develop a joint distribution of mean wind
speed and wave height. Haver and Winterstein (1990) suggest modelling mean wind
speed conditional on the extreme significant wave height (or extreme wave height) as
a normal variable with a constant standard deviation and a mean varying linearly with
the actual significant wave height (or extreme wave height) as p.,=A+BH1. In a
more refined analysis the coefficients A and B can also be treated as basic random
variables to reflect the uncertainty in relating wind speed to wave height.

The instantaneous wind speed, which is used in determining environmental forces, will
be different from mean wind speed due to the effect of wind gust. This effect is
usually accounted by treating instantaneous wind speed as V=V 0(1+G), where G is
the gust factor and V,,0 is the mean wind speed. The gust factor can be modelled as
a normal variable with a mean of zero and standard deviation a 0 obtained from
measurements. In order to account for measurement and statistical uncertainties a0
itself can be modelled as an independent normal random variable. Haver and
Winterstein give values of 0.03 and 0.005 for the mean and standard deviation of a02.

Current S peed V:
Current can be either a tidal current or wind generated current or a combination of
both and normally varies with depth below the water surface. It is often the practice
to measure only the surface value of current and adopt a current profile to describe
current variation with depth. Uniform, triangular, bi-linear and power-law current
profiles axe commonly used. In a probabilistic analysis, therefore, it is convenient to
model the surface current V as a random variable and assume the current profile to
be deterministic. Again where data are available it is preferable to model current
distribution conditional on wave height. A normal distribution has been suggested by
Haver and Winterstein (1990) to model current distribution conditional on extreme
significant wave height. The conditional mean has been assumed to vary linearly with
extreme significant wave height as p=A-I-BH while the standard deviation is
5-8
assumed to be constant. In a more refined analysis uncertainty in p can be modelled
by treating A and B as random variables.

When enough data are not available to develop joint/conditional distribution of all
environmental variables, each of the variables can be specified a marginal distribution
based on available data and correlations can be specified between pairs of variables
based on subjective judgement. The modelling of parametric uncertainty (i.e.
uncertainty in distribution parameters) will also depend on availability of data. When
the available data is poor it may be preferable to subjectively increase the variability
(for example increasing the coefficient of variation) to account for statistical and
model uncertainties.

5.2.2 Calculation of Hydrodynamic Loading:

Calculation of hydrodynamic loading due to waves and current on a jacket structure


requires additional parameters such as thickness of marine growth, drag and inertia
coefficients in Morison's equation all of which are subject to significant uncertainty.
In addition, the assumption of regular and long-crested waves, the use of simplified
wave theories, such as Airy or Stoke's , the empirical nature of Morison's equation
etc can cause considerable uncertainties in the computed wave force on the structure.
Treatment of these sources of uncertainties are discussed in the following in the
context of deterministic, discrete wave static response analysis.

Thickness of Marine Growth tmgi


Marine fouling is a biological process caused by the build up of marine species such
as mussels, kelps, sea anemones, barnacles etc. The thickness of marine growth
increases with time controlled by such factors as water depth, temperature, oxygen
content etc. and generally decreases in thickness from the sea surface to the mud-line.
It is a common practice to report marine growth thickness measurements in two or
three different depth zones and the fouling is cleaned periodically.
5-9
Marine fouling affects hydrodynamic loading on a structure in two important ways.
First, it increases the effective diameter of the jacket members, leading to increased
fluid loading. Second, the surface of members becomes more rough, leading to
increased drag coefficient, and hence increased hydrodynamic loading. The second
effect is considered in connection with the uncertainty in drag and inertia coefficients
below. Since the marine growth thickness on any member varies, depending on its
location below the surface and the state of cleaning of the structure it is likely that
marine growth on different members will not be fully correlated. However, as a
reasonable simplification it can be assumed to be fully correlated all over the structure.
In a probabilistic analysis it is thus convenient to use a substitute variable Bmg with
unit mean which is multiplied with the stipulated marine growth at any depth to obtain
the actual marine growth thickness. A lognormal distribution is often used to model
Bmg see for example Olufsen and Bea (1990). Coefficient of variation in marine
growth thickness can be expected to be of the order of 0.40-0.50.

Dra2 Coefficient Cd:


In deterministic design practice it is common to use a constant value, typically 0.7,
for all members in a structure. For this reason in most studies on reliability of
offshore structures, the drag coefficient is modelled as a single random variable, fully
correlated for all members in a structure. The variability in drag coefficient arises due
to many sources such as surface roughness of members, state of flow, presence of
current, etc. Even laboratory tests on smooth cylinders show considerable scatter in
measured drag coefficients. The drag coefficient can be modelled as a lognormal
variable, fully correlated for all members in the structure. A mean value of 0.70 and
a standard deviation of as high as 0.25, see for example Turner and Baker (1988).

The assumption of full correlation in Cd values for all members in a structure is


unrealistic. But at the present enough data are not available to develop a correlation
structure for Cd values of different members. One of the alternatives is to split the
variability in Cd in to two components, viz. (1) one arising due to the inherent
variability in Cd values as observed in laboratory tests on smooth cylinders, and (2)
the other arising due to dependence of Cd on such factors as surface roughness, state
5-10
of flow, current, etc which are themselves uncertain quantities. This approach has
been used by Carr and Birkinshaw (1988, UT! 537), in which the inherent variability
in Cd is modelled by a multiplicative random variable B with a mean value of unity
and a coefficient of variation of 0.15. The mean value of Cd is obtained as a function
of surface roughness, Keulegan-Carpenter number and current speed. The functional
relations between these quantities have been developed mainly from the results of
OTS experiment and are given in the above reference. This approach helps to
compute mean Cd values more accurately for different members in a structure instead
of using a single value for all members. At the same time a more realistic correlation
between Cd values of different members is automatically built-up by the functional
dependence on other factors.

Inertia Coefficient C.,


Similar to drag coefficient, the inertia coefficient is often modelled as a fully
correlated random variable in most studies on system reliability. A normal distribution
with a mean value of 1.80 and a standard deviation of 0.27 has been used by Baker
and Ramachandran (1981). The approach used by Carr and Birkinshaw is to model
the inherent variability in Cm through a multiplicative random variable Bm which is
lognormally distributed with a mean value of 1.05 and coefficient of variation of 0.34.
The mean value of Cm is computed as a function of surface roughness which itself is
modelled as a random variable with a uniform distribution in the range lOmm-25mm.

Model Uncertainty in Hvdrodvnamic Force Calculation:


In addition to the variability in design quantities mentioned above, significant
uncertainties can arise due to the use of a particular wave theory, such Stoke's theory,
the assumption of long-crested regular waves, ignoring shielding effects and the use
of the empirical Morison formula. The quantification of these uncertainties is
generally difficult. Bea et al (1988) have estimated the bias (B F) related to the
extreme base shear predicted from offshore measurements versus that using a discrete
wave approach with Stoke's fifth-order wave theory. Assuming uni-directional waves
the mean and standard deviation of BF were estimated to be 1 22 and 0.23,
respectively For multi-directional waves the corresponding values were found to be
5-11

0.88 and 0.30. Since constant values of Cd and Cm were used in the analysis the
statistics of B F represents the combined uncertainty in wave kinematics, drag and
inertia coefficients and hydrodynamic force calculation. This approach automatically
incorporates all the correlation effects of drag and inertia coefficients between
different members in a structure. However, more studies on similar lines with on-site
measurements are required to develop credible statistics for BF.

5.2.3 Global Structural Analysis:

Additional sources of uncertainty arise in the structural analysis to compute nodal


displacements and forces and member stresses. These are due to finite element
idealisation of the structure, uncertainty in joint flexibility, non-linear
soil-pile-structure interaction behaviour etc. Uncertainty due to these factors is
believed to be relatively small compared to the uncertainty in the calculation of
hydrodynamic loading. Statistics for this model uncertainty can be developed in a
similar manner discussed above by comparing measured stresses with those computed
using a standard structural analysis method. It is often convenient to model the
combined uncertainty in wave kinematics, hydrodynamic loading and structural
analysis with a single variable Bga representing the uncertainty in member nominal
stresses computed from a global structural analysis. From the information given in
Wirsching (1984) it is reasonable to assume Bga to be lognormally distributed with a
mean value of 0.8 and a coefficient of variation of 0.30. This variable also
incorporates the variability in drag and inertia coefficients.

5.3 UNCERTAINTIES IN FATIGUE LOADING

For the fatigue design of North Sea jacket structures a spectral method is commonly
employed. In this method uncertainties can arise in the modelling of ocean waves
both in the short-term and long-term contexts. Additionally, significant uncertainties
will be involved in the calculation of transfer functions between sea-surface elevation
5-12
spectrum and response spectra of nominal stresses in various brace members. As
discussed in the foregoing section on extreme wave load modelling each source of
significant uncertainty should preferably be modelled through an explicit basic
variable. However, in the case of cumulative fatigue loading uncertainties in the
calculation of transfer functions tend to dominate when compared to the uncertainties
in the modelling of wave climate. For this reason and considering the computational
cost of including full spectral response analysis within an iterative reliability analysis
simpler methods of uncertainty modelling may be desirable in some situations. In the
following three alternative forms of uncertainty modelling for fatigue loading are
discussed. The suitability of each of these methods should be assessed depending on
the application.

5.3.1 Explicit Modelling of All Loading Variables:

This is the most rigorous approach for modelling of uncertainties in fatigue loading
and should be preferred when adequate data are available for modelling individual
sources of uncertainty and computational time is not a restriction. This method can
also be used to calibrate simpler methods discussed in the later sections. The various
sources of uncertainty in modelling fatigue loading can be categorised into three main
groups and accordingly the treatment of these uncertainties is discussed under the
following three heads.

5.3.1.1 Long-term wave climate modelling:

As discussed in chapter 3, a long-term wave climate description consists of a joint


distribution of significant wave height (I-I s ) and zero-crossing period (T1) as these are
used as the parameters of a short-term model. In addition, a probability mass function
giving the probabilities Df each of the main wave approach directions considered ii
the analysis. Uncertainties can arise in both the descriptions.
5-13
The joint distribution of significant wave height and zero-crossing period is usually
expressed in the form of a scatter diagram based on recorded data at a particular
location. This can also be represented in terms of a marginal distribution for
significant wave height and a distribution for zero-crossing period conditional on the
significant wave height, see chapter 3 for a detailed discussion. The marginal
distribution of significant wave height is often modelled as Weibull although some
researchers use a combination of lognormal (for low sea-states) and Weibull (for high
sea-states) distributions, see for e.g. Haver and Natwig (1991). The conditional
distribution of zero-crossing period is often modelled as lognormal.

The uncertainty in the long-term description of wave climate arises mainly due to
statistical reasons. The data on wave height/period are usually obtained over a limited
duration which may not adequately account for the seasonal and yearly climatological
variations. Even data extending over 3-5 years may not give a good long-term
description. Data for the chosen site are sometimes not available at all and data from
neighbouring areas have to be used. In order to account for these uncertainties in a
reliability analysis the distribution parameters describing the significant wave height
and zero-crossing period have to be modelled as basic random variables. However,
since most of the fatigue damage is caused by waves of moderate amplitude, for
which a reasonably large sample of data is generally available, the statistical
uncertainty is generally expected to be small. For example, Kirkemo (1988) has
modelled the bias in E[H] as a lognormal variable with a mean of unity and
coefficient of variation of 0.1 while the bias in a[HJ is modelled with a mean of unity
and coefficient of variation of 0.1. Similarly, the bias in E [TJ and a[T] are each
modelled as lognormal variables with a mean of unity and coefficient of variation
of 0.06. These values are, however, based on judgement and not on statistical analysis
of environmental data at any location.

The probability mass function of wave approach direction is usually obtained from
wind measurements as directional wave data are often not available. However, in
sheltered sea environments, such as the North Sea, directional distributions based on
wind measurements can be quite different from those obtained from wave data. Haver
5-14
and Natwig (1991) have examined this problem for a northern North Sea location. A
deterministic assessment of fatigue damage using directional distributions obtained
from wind and wave measurements gave very different results and it was found that
computed fatigue damage was very sensitive to the selected directional distribution.
In a reliability analysis, the specified directional probability values can be randomised
by using a multiplicative random variable. Since data are not available, this variable
can be modelled approximately by a normal distribution with a mean of unity and a
coefficient of variation in the order of 0.10-0.20. Because of lack of data it often
becomes necessary to assume that wave direction probability is statistically
independent of significant wave height and zero-crossing period.

5.3.1.2 Short-term Wave Climate Modelling:

In the short-term, the sea-surface elevation is often modelled as a zero-mean,


stationary Gaussian process, see chapter 3. Thus the sea-state is completely
characterized by a directional wave spectrum, which is written as the product of the
uni-directional spectrum and a directional spreading function, see chapter 3 for details.
The sea-state spectrum used in practice, such as Pierson-Moskowitz, JONSWAP etc.,
is an empirical model the parameters of which are related to the significant wave
height and zero-crossing period which characterize the sea-state. Thus the short-term
wave model is defined conditionally on the significant wave height and zero-crossing
period which have a long-term fluctuation.

Uncertainty in the above model can arise from the choice of a particular spectral
model, such as P-M or JONSWAP. Once a choice has been made, additional
uncertainty may arise from the assumed form of the spectrum, i.e. the choice of
exponents and in the case of a P-M spectrum and the peak enhancement factor
in the case of a JONSWAP spectrum, refer Eq.3.i in chapter 3IiiãIibilit
analysis, uncertainty m short-term wave modelling can be accounted by considering
and , and additionally in the case of JONSWAP spectrum, as random variables
For example, Kirkemo (1988) has used a P-M spectrum with as deterministic while
5-15
as a normal variable with a mean of 4 and coefficient of variation of 0.05.
However, the reliability mdcx was found to be relatively insensitive to the uncertainty
in . More recently, Haver and Natvig (1991) have studied deterministically the
sensitivity of fatigue damage in a TLP structure to different forms of JONSWAP
spectrum. It was found that the use of a JONSWAP spectrum as against a P-M
spectrum results in an increase in fatigue damage by about 5%. However, the
computed fatigue damage was found to be less sensitive to the form of the spectrum.

The modelling of directional spreading can also introduce additional uncertainties.


This can be accounted by treating one of the coefficients in the adopted spreading
function as random. When relevant data are not available this variable can be
modelled using a normal distribution with a coefficient of variation of about 0.05.
From a deterministic study Haver and Natvig (1991) have found that the form of the
spreading function is not very crucial to fatigue damage. However, neglecting
directional spreading can have a significant influence. The effect is seen to be
enormous when all sea-states are assumed to approach from the most unfavourable
direction. However, when the sea-states are distributed into several sectors in a
realistic manner, the effect of short-crestedness is significantly reduced. In this case,
accounting for short-crested waves result in a decrease in computed fatigue damage
typically in the range of 5%-15%.

In addition to the above, uncertainty can also arise from the Gaussian assumption for
sea-surface elevation but this is usually a problem for shallow-water structures. The
uncertainties in the modelling of wind and current conditions is particularly not crucial
in fatigue assessments as these contribute only to the mean stress level for fatigue
cycling which has only a secondary influence on fatigue damage.

5.3.1.3 Response Transfer Function:

The spectral response analysis method has already been presented in detail in
section 3.3.2 of chapter 3. The transfer function for member nominal stresses is
5-16
developed in broadly three steps: (i) transfer function to calculate particle kinematics
using a linear Airy wave theory, (ii) transfer function to calculate hydrodynamic forces
on members using a linearized Morison equation and (iii) transfer function for member
nominal stresses using stiffness and modal analysis of the structure. Significant
uncertainties can arise in each of these steps and can be treated as discussed in the
following. Additional uncertainties also arise in the calculation of stress concentration
factors for determining the spectrum of hot-spot stress at a tubular joint and these are
considered in the next section.

As discussed in section 5.2.2 in the context of extreme load modelling, considerable


uncertainties are involved in the calculation of hydrodynamic loading due to uncertain
extent of marine growth, uncertainty about the values of Morison coefficients C 4 and
Cm and these can be treated in a similar manner as discussed in section 5.2.2. In
addition, the use of a linear Airy wave theory for particle kinematics and linearization
of the drag term in the Morison equation contribute to significant model uncertainties
in the transfer function.

The model uncertainty in hydrodynamic force calculations can be treated in a similar


manner as discussed in section 5.2.2. It is seen that the use of a spectral approach
based on Airy wave theory results in a bias in the computed wave force. The ratio
of horizontal base shear and moment computed by a discrete wave approach using a
Stoke's wave theory to that computed by a spectral approach, bF, can be shown to be
of the order of 1.10-1.30, with higher values for more severe sea-states, see Olufsen
and Bea (1990). Thus the mean value of the total bias in stochastic wave force
calculation can be calculated as a product of PBF and PbF• But the standard deviation
of the model uncertainty is found to depend on the significant wave height for a
particular sea-state, generally decreasing for more severe sea-states.

Additional model uncertainties will also be involved injhe stiffnessLmodal analysisJo


determine transfer functions for member nominal stresses. As discussed in
section 5.2.3, these uncertainties can be combined with the model uncertainty in the
calculation of particle kinematics and hydrodynamic loading and expressed through
5-17

a single variable Bga. Nerzic and Lebas (1988) have determined the statistics for this
variable by comparing measured stresses in brace members in an actual offshore
structure to that computed using a standard spectral procedure. A lognormal model
with a mean value of 0.88 is found to fit the data well. The coefficient of variation
of Bga was found to vary in the range of 0.20-0.40, generally increasing for sea-states
with lower significant wave height. These values are comparable with those used by
Wirsching (1984). Note that the variability in Bga also includes the variability in Cd
and Cm coefficients.

Additional uncertainties will be involved in the use of an empirical short-term density


function for stress-ranges to compute cumulative fatigue loading. However, from the
comparison of various commonly used density functions, presented in section 3.5.1 of
chapter 3, it can be seen that the choice of a particular model only contributes to the
bias in cumulative loading rather than to uncertainty. If a long-term model is used,
as described in section 3.5.3, to fit a single density function to the stress cycles from
all sea-states additional uncertainty will arise due to fitting errors. This is considered
in the next section.

5.3.2 Uncertainties Modelled Through the Parameters of the Long-term Mode):

Treatment of uncertainties in fatigue loading through an explicit basic variable for


each significant source of uncertainty, as presented above, although provides a more
accurate description of uncertainty, this approach will require the entire spectral
analysis procedure to be repeated several times within a iterative reliability analysis.
This will be prohibitively expensive, especially in a system reliability analysis under
fatigue conditions, where a large number of failure sequences have to be analyzed.
In such situations it is better to express the uncertainty in fatigue loading through the
parameters of the long-term Weibull model developed for individual joints. These
variables can then be used within a reliability analysis without the need to repeat the
full spectral calculations. A procedure for developing uncertainty statistics for a
2-parametered long-term Weibull model of stress-ranges from the uncertainty models
5-18

for all the individual loading variables has been proposed by Skjong and Torhaug
(1991). The method involves the use of the FORM iterative technique to compute
long-term probabilities corresponding to few specified stress-range levels and fitting
a Weibull model to the results. This approach can also be used for the 5-parametered
long-term Weibull model for stress-ranges proposed in section 3.5.3.

5.3.3 Uncertainties Modelled Through a Single Variable:

From the discussion in section 5.3.2 where the individual sources of uncertainties in
fatigue loading are discussed, it is clear that uncertainties in response transfer function
are considerably higher compared to those in environmental description. Since the
main contribution to the fatigue damage is from large number of low and medium
waves and not from the less frequently occurring extreme waves, the prediction of the
sea-state statistics is not the major source of uncertainty, Baker (1985). This inference
can in fact be verified from the results of an example reliability analysis presented in
Kirkemo (1988). The results show that the probability of failure is less sensitive to
the uncertainty in significant wave height, zero-crossing period and the variables
describing the form of sea-surface elevation spectrum, than to other variables.

In view of the above, all the sources of uncertainty in fatigue loading can be
conveniently expressed through a single variable. The model uncertainty variable Bga
which accounts for the uncertainty in response transfer function can be used for this
purpose. As discussed before the statistics for this variable should be developed from
comparisons with on-site measurements. This uncertainty can be slightly augmented
to include the uncertainties in environmental description. The contribution of
uncertainty in environmental description to the total uncertainty in fatigue loading is
expected to be around 10%. From the limited information available at present Bga can
be modelled as a lognonnal variable with a mean value of ityandcoefficient of
variation of 0.25-0.50. This variable can be reasonably assumed to be fully correlated
for all members in the structure. This provides a simplified model]ing of uncertainties
in fatigue loading. The advantage of this approach is that the uncertainty in fatigue
5-19
loading can now be applied directly to the computed member nominal stresses by
multiplying with the model uncertainty factor Bga. This avoids the need to include the
full spectral calculations within the iterative reliability analysis, thus saving
considerable computational effort.

5.4 UNCERTAINTIES IN STRENGTH PREDICTION MODELS

In addition to the uncertainties in the modelling of loading and response of an offshore


structure, uncertainties will also be involved in the models used for strength prediction
of primary members and tubular joints in the structure. A number of models for
estimation of plastic collapse loads, stress analysis, fatigue crack propagation and
fracture assessment of tubular joints have been presented in chapter 4 and the
treatment of uncertainties involved in these models is discussed in the following.

5.4.1 Plastic Collapse Load Model:

Plastic collapse load models have been developed primarily from tests on scaled steel
tubular joints under individual modes of loading. Very limited test results are
available to assess the effect of presence of chord loads and combined loading in the
brace member. There is also no information about the strength of multi-planar and
stiffened joints. All the tests were invariably based on single joints isolated from the
rest of the structure. Because of these reasons, combined with the inherent scatter in
test results, the use of the empirical models in predicting the behaviour of real tubular
joints in an offshore structure will be subject to considerable uncertainty.

A systematic study to quantify the uncertainty associated with these models is not
possible because of the lack of homogeneous sets of test data and the lack of
information about the performance of real joints in an offshore structure. The
statistical values reported in Dept. of Energy (1989b) for each mode of loading are
shown in Table 5.1 and they show the extent of scatter involved in test results.
5-20

Table 5.1: Scatter in static strength test results on tubular joints

Loading Mode No. of Tests Mean COY

Axial Compression 42 1.02 0.30


Axial Tension 18 1.00 0.40
In-plane Bending 31 0.98 0.35
Out-of-pI. Bending 14 1.00 0.30
All Load Cases 105 1.00 0.34

The apparently large scatter in test results arises due to mixing of data-sets obtained
from across the world from tests conducted under different conditions (for example
fixed or pinned supports, load or deflection control etc). The scatter also includes the
effect of inherent randomness in material yield strength, weld geometry etc. The
difference in coefficient of variation for data-sets with different modes of loading may
be largely due to different number of test results in each category and the source of
data. In view of these it is suggested that the model uncertainty in collapse load
prediction be accounted through a single variable B 01 which gives the ratio of the true
collapse load of a joint to the load predicted using the model. Since the proposed
model is based on mean value test results the mean value of B 1 is expected to be
unity. The coefficient of variation obtained from a homogeneous data-set of 12 tests
under axial compression reported in Dept. of Energy (1989a) is found to be about
0.15. Augmenting this variability with the uncertainty in other factors in estimating
the performance of real tubular joints, the variable B 1 is expected to have a
coefficient of variation of 0.15-0.25. A lognormal distribution is found to give a good
fit to the test results. The variable B 1 may be assumed to be fully independent for
different joints, as the data is mainly obtained from test results on individual joints.
5-21
5.4.2 Stress Analysis Model:

Models for stress analysis of tubular joints have been presented in chapter 4 where it
was suggested that simplified empirical models could be conveniently used in a
reliability analysis. A number of parametric formulae, such as Kuang,
Wordsworth-Smedley, UEG, Efthymiou etc, are available for estimating stress
concentration factors (SCFs) for the geometric stress at critical points around the
tubular joint. Comparison of SCFs predicted by these formulae with test results using
steel models show considerable scatter, UEG (1985). The discrepancy between
predicted and measured SCF values can be expressed in terms of the factor B giving
the ratio of measured SCF to that predicted using one of the parametric formulae.

Statistical analysis for the model uncertainty factor B using the formulae referred
to above derived from comparison with measurements on steel models has been
reported in Lalani et al (1986) and Dover et al (1991). Table 5.2 gives an idea of the
level of scatter inherent in the use of various parametric formulae. Although the
statistics vary slightly for different joint types, loading mode and location of hot-spot
it is considerably simpler if a single variable is used for all load cases, the statistics
of which may be taken as in the last row of Table 5.2. The factor B can be
modelled as a lognormal variable, independent between different joints in a structure.

Table 5.2: Comparison of Model Uncertainty in SCF Formulae

Kuang (1975) UEG (1985) Efthymiou (1988)


Loading Mode
Mean COV Mean COV Mean COY

Axial Loading 0.82 0.28 0.74 0.35 1.0 0.22


In-pi. Bending 0.90 0.30 0.74 0.30 0.82 0.30
Out-of-pi. Bend 0.84 0.45 0.66 0.40 0.80 0.30
All Load Cases 0.85 0.33 0.75 0.35 0.98 0.28
5-22
The above SCF formulae are based on tests on simple tubular joints. The use of these
formulae for multi-planar joints as in real structures may lead to additional
uncertainty. The methods used to interpolate SCF values between saddle and crown
points and the combination of stresses due to individual loading modes to compute
total hot-spot stress at a location will also add some uncertainty. These additional
uncertainties are, however, more difficult to quantify due to lack of data.

A fracture mechanics approach requires degree-of-bending factors (DoBs) in addition


to SCFs to compute through-thickness stress distribution. The use of parametric
formulae, such as those given by Conolly et al (1989), or finite element methods will
involve significant uncertainty. The model uncertainty in degree-of-bending factors
can be similarly expressed through a factor BdOb which gives the ratio of measured
DoB to that computed using a model. The data given in Conolly et al (1989) for their
parametric formulae suggest a mean value of 0.85 and a coefficient of variation of
0.15. The factor BdOb can be modelled as a lognormal variable statistically
independent of Bf and independent between joints.

5.4.3 Fatigue Crack Propagation Model:

The fatigue crack propagation model presented in chapter 4 is likely to be subject to


considerable uncertainty. The uncertainties arise from the use of simplified empirical
formulations for stress intensity factors which are basically developed for flat plates,
unknown distribution of residual stresses, modelling of weld geometry effect, crack
coalescence, relaxation of residual stresses and load shedding during crack growth.
It is difficult to quantify the uncertainty for each of these sources separately due to
scarcity of data. It is convenient to express all the above sources of uncertainty
through a single basic variable B1 which gives the ratio of stress intensity geometry
correction factor obtained by experiment, Y,, to that computed using the proposed
model This is because the major sources of uncertainty referred above all
pertain to the uncertainty in the calculation of stress intensity factors for tubular joint
cracks. The statistics for this variable can be developed by comparison with
5-23

experimental compliance function curves. The statistics are expected to vary


depending on such factors as crack depth, joint type and loading mode but it is
convenient to use a single variable with statistics representative of all cases. In view
of the lack of sufficient data at present it is proposed to model the factor B f as a
lognormal variable with a mean of unity and coefficient of variation in the range of
0.15-0.25. Lower variability may be used if the stress intensity factors are computed
using finite element models such as line-spring and weight function models. The
variable B51f is assumed to be statistically independent of B5f and BdOb and
independent between different joints in a structure.

5.4.4 Fracture Assessment Model:

The major sources of uncertainty in the fracture assessment model proposed in


chapter 4 include: uncertainty in the failure assessment curve, the calculation of plastic
collapse parameter Lr arid the calculation of fracture parameter Kr The latter tWO
sources can be largely accounted for through the variables B and B f discussed
previously. The uncertainty in the failure assessment curve arises from the fact that
a universal curve is used regardless of the material type and geometry and the true
strain-hardening behaviour is not accounted for.

The CEGB R6 universal failure assessment curve has been compared with more
accurate solutions using J-integral calculations and also with fracture tests on
components of different material and geometry, see Ainsworth et al (1986). The
results show that the use of the Option 2 failure assessment curve with material
specific strain-hardening gives a good estimate of fracture initiation load. The use of
the universal failure assessment curve (Option 1) gives a conservative estimate in
comparison with Option 2 curve. The level of uncertainty is seen to be less when
compared to that involved in the calculation of L and K parameters for tubular joints.
Since comparisons with tubular joints are not available the uncertainty in failure
assessment curve is ignored in comparison with B and B
5-24
5.5 UNCERTAINTY IN MATERIAL PROPERTIES

The material properties, such as crack growth parameters and fracture toughness, used
in the assessment of tubular joints are inherently random in nature. The uncertainty
in material properties should be explicitly accounted for in a reliability assessment.
The uncertainty in material properties is made up of both inherent variability and
statistical uncertainty. With the available data, although reliable statistics can be
developed for variability in individual material properties, it is rather difficult to
quantify the correlation in material properties between various joints in a structure.
The correlation in material properties between joints has a significant influence on the
computed system reliability and considerable care and judgement is required in
modelling them. The treatment of uncertainties in various material properties required
in the reliability assessment of tubular joints is discussed in the following.

5.5.1 Plastic Material Properties:

The plastic material properties of relevance in the assessment of tubular joints are the
yield stress, the flow stress, the ultimate tensile stress and if available the full
strain-hardening curve. Of these yield stress is the only significant variable for the
strength prediction models proposed in chapter 3 and uncertainty modelling for this
is considered here.

The variability in yield stress is made up of many sources such as, manufacturer, plate
thickness, within cast variability and variability from one cast to the other,
Baker (1972). The probability distribution for yield stress should be built-up from the
knowledge of the conditional distributions for each of these sources. When such
detailed information is not available a lognormal distribution is usually used for yield
stress considering a fibre bundle model for the plate strength. For high strength steels,
of grade BS 4360 50D, usually used for stub and can sections of tubular joints, a
mean value of 380 N/mm2 and a standard deviation of 38.0-48.0 N/mm 2 may be used
for modelling the variability in yield stress.
5-25

5.5.2 Fatigue Material Properties:

In the S-N approach the fatigue resistance of a joint is expressed in terms of an S-N
curve while in the fracture mechanics approach it is expressed in the form of the
da/dn vs K curve. The uncertainty modelling for both the cases is considered here.

S-N Curve Parameters: S-N curves are developed from fatigue tests on tubular
joints and generally a very high scatter is observed in the number of cycles to failure.
The uncertainty in S-N curve can be expressed through the parameters m and K in
Eq.4. 11. Because of the mathematical form of the equation the parameters m and K
are expected to be highly correlated and it is common to consider only one of the
parameters as random and the other as fixed. Considering K as random the statistic
for this variable can be obtained from the test data on number of cycles to failure.
The variable is often modelled as lognormal, see for example Wirsching (1984). For
the test data obtained under the UKOSRP programme which form the basis of the
Dept. of Energy (1984) 'T' curve, the mean value and standard deviation of K for
16mm specimens are found to be l.012E+13 and 5.850E+12 with the stress-range
expressed in N/mm2, see Thorpe and Sharp (1990). If the S-N curve has an endurance
limit, this also can be modelled as a random variable. When a bi-linear S-N curve is
used, the parameter K2 can be treated as a random variable, correlated with K 1 , or
made functionally dependant on K 1 . Similarly the uncertainty in thickness correction
can be considered by modelling the exponent , in Eq.4.12 as a lognormal random
variable with a mean of 0.3 and a standard deviation of 0.03.

Miner's Sum: The Miner's hypothesis of linear damage accumulation states that the
component will fail by fatigue when the accumulated damage D exceeds unity.
However, a number of tests under variable amplitude loading indicate a scatter in the
value of the accumulated damage at failure. This could be attributed as a model
uncertainty in the Miner's hypothesis. This uncertainty can be modelled by treating
the Miner's damage sum at failure (A) as a random variable. Wirsching (1984)
suggest that A be modelled as a lognormal variable with a mean of unity and
coefficient of variation of 0.3.
5-26

Crack Growth Parameters: Crack propagation rate data are usually obtained from
tests on simple specimens and exhibit considerable scatter. The uncertainty modelling
for crack growth depends on the crack propagation model used. As reviewed in
section 4.4.3.1 of chapter 4, a number of probabilistic crack propagation models
attempt to describe the scatter in test results by treating crack propagation as a
stochastic process, for e g. a lognormal process or a Markov process. However, when
a simple Paris type crack propagation model is used, as proposed in section 4.4.4, a
random variable description is usually adequate. In this approach, the crack growth
parameters, C and m, in the Paris law can be used to describe the scatter in crack
propagation data.

The crack growth parameters C and m usually exhibit a high negative correlation and
it is therefore common to use a fixed value of m and express all the uncertainty
through C. The variable C is well described by a lognormal model and for the
BS 4360 grade steel a mean value of 4.5E-12 (with m=3.3, da/dn in rn/cycle and EK
in MPairrn units) has been suggested by Aaghaakouchak et al (1989). Greatly
differing values have been suggested by different workers for the coefficient of
variation in C, see Kirkemo (1988), but a value between 0.30-0.40 seems to be
reasonable. Systematic data collection and statistical analysis is required for this
variable for typical offshore construction which should account for the variation in
material properties of the weld metal, HAZ and parent plate. Similarly the uncertainty
in fatigue threshold can be accounted by treating the threshold stress intensity factor
range LKth as a random variable.

Initial Defect Size: Crack like defects are invariably present in welded joints and
fatigue cracks often initiate from such defects. The weld defects are of several types
and defect size and occurrence rates are found to be highly random and being
influenced by such factors as fabrication yard, welding procedure, welding position,
type of joint etc. Baker et al (1988). Very limited data have been reported on weld
defect size distribution typical of offshore construction and tubular joints. Some of
the reported studies on plated joints have been reviewed by Kirkemo (1988). Recently
statistical analysis of a large amount of weld defect data obtained from the Conoco
5-27

Hutton TLP structure has been carried out by Kountouris and Baker (1989). For use
in a reliability analysis, the distribution of defect sizes should be based on defects
existing in a structure entering into service, considered acceptable according to quality
control standards, as well as those remaining undetected during fabrication. Thus
defect occurrence rate, the amount and quality of NDE used should be taken into
account in developing distribution for weld defect size. From the limited data
available the mean value and standard deviation of weld defect depth a 0 can be
estimated to be 0.15mm and 0.10mm for "sound" quality welds. Lognormal,
exponential and Weibull distributions have been used by researchers to fit weld defect
data. Similarly, initial defect aspect ratio, defined as the ratio of initial defect depth
to defect semi-length (ajc 0 ) may also be modelled as a lognormal variable with a
mean of 0.62 and coefficient of variation of 0.4, Kountouris and Baker (1989).

5.5.3 Fracture Material Properties:

The fracture resistance of welded joints is influenced by material parameters such as


fracture toughness and residual stresses. Uncertainty modelling for these variables is
discussed below.

Residual Stresses: Residual stresses are introduced into tubular joints due to welding
and fabrication process and are a subject of considerable uncertainty. Systematic
studies on statistical modelling of residual stresses in tubular joints have not been
reported in the literature because of the enormous difficulty in measuring residual
stresses. The limited measurements by Porter-Goff et al (1988) indicate that the
magnitude of residual stress can vary considerably often reaching yield value at the
surface. Based on these measurements residual stress at the surface of the plate in a
tubular joint can be modelled as a lognormal variable with a mean value of 300 MPa
(tension) and a standard deviation of 75 MPa. The through-thickness distribution and
residual stress relaxation during crack growth may considered deterministic and
modelled as discussed in chapters 3 and 4.
5-28

Fracture Toughness: The fracture toughness of the material for use in the fracture
assessment methodology proposed in section 4.5.2 may be in terms of the
linear-elastic fracture toughness 1Siiat or the crack tip opening displacement Data
on fracture toughness is usually obtained from tests on simple specimens and generally
a large scatter is observed in test results. In addition, the actual fracture toughness of
tubular joint in a structure may be affected by a number of factors such as, plate
thickness, stress state, mode of loading, service temperature etc. The influence of
these should be considered in developing probability distributions for fracture
toughness. Data should be collected and analyzed separately for the weld metal, HAZ
and the parent plate as they exhibit different material micro-structure and significant
difference in toughness values. From the limited data reported in the literature for BS
4360 50D steel fracture toughness for the HAZ material may be modelled as a
lognormal variable with a mean value in the range of 90-120 MPaIm and a coefficient
of variation of 0.20-0.30 MPaJm.
5-29
5.6 RELIABILITY ANALYSIS FOR THE LIMIT-STATE OF FATIGUE

In a deterministic approach for design against fatigue failure of tubular joints, see for
example Dept. of Energy (1990), all the loading parameters are taken at their expected
values while the resistance parameters are taken at values corresponding to mean
minus two standard deviations. Additional safety is ensured by keeping the computed
fatigue lives by a factor of 2-10 times higher than the planned service life of the
structure. However, in view of the considerably large sources of uncertainty, both in
the loading and fatigue resistance parameters, a fully probabilistic approach to fatigue
design is considered appropriate. In this section reliability analysis methodology for
the fatigue limit-state is presented both using an S-N approach and a fracture
mechanics approach. The sensitivity of the probability of failure of a tubular joint to
uncertainties in various parameters is studied using a number of examples.

5.6.1 Reliability Analysis using the S-N Approach:

Using the expression for fatigue damage, Eq.4. 16, from chapter 4 and the expression
for cumulative loading, Eq.3.31, from chapter 3, the damage due to ; cycles of
loading in a service life of t can be computed as


DL - BIJB,CJE[SA] (5.1)

where i is the long-term average frequency of stress cycles, computed using Eq.3.52,
and the mth expected value of stress-range can be computed as in Eq.3.41 using the
short-term density function model or as in Eq.3.53 using the long-term density
function model presented in chapter 3. The variables and are introduced to
account for the uncertainty in the calculation of the global transfer function and the
stress concentration factors, respectively. The uncertainty in other loading variables,
such as sea-state parameters, marine growth, drag and inertia coefficients, etc can be
introduced implicitly by making aIE[Sm] a function of these random variables. When
5-30

a multi-linear S-N curve is used the damage due to stress cycles falling in different
segments can be calculated separately and added together.

In addition, considering the value of the Miner's damage sum at failure A as a random
variable the probability of failure due to fatigue can be expressed as

pf - P[L^DJJ
(5.2)

Two approaches are presented in the following for the computation of pg

Lo2normal Formulation:
This formulation has been extensively used within the API PRAC project, see
Wirsching (1984), for the derivation of allowable stresses for fatigue design based on
reliability considerations. In this approach, only A, K, BStf and BSCf are treated as
random variables all of which are assumed to follow a lognorrnal distribution. With
this an expression for the reliability index can be written as

_______
13 -

(53)
)m2]
- ln [(1 . V,)(1 +V)(1 ^V )m2(1 i-V
'4

where V denotes the coefficient of variation and the tilde denotes the median value
of a variable. The median fatigue damage can be calculated from Eq.5. 1 using median
values for the random variables K, Bgtf and B51. The probability of failure can be
obtained from the reliability index using p 1=cI)(-f3) as discussed in chapter 2.

The main advantage of this approach is that an exact closed form expression of the
reliability index can be obtained which makes reliability analysis simple and efficient.
However, the main disadvantage is that all the variables have to be assumed to follow
a lognormal distribution. Moreover, the method cannot be used if an explicit
expression for damage in terms of all the basic variables is not available. Thus the
uncertainty in other loading variables, thickness correction, endurance limit, etc cannot
be included in a reliability analysis.
5-31
The sensitivity of the reliability index with respect to the coefficient of variation of
each of the variables can also be derived analytically as

a - V1%

V,
- -

v z: i+v,
(5.4)
m2V2,,
____ -
a, 1+V

m2V8
____ -

v B T1112

Generalized Formulation:
In this approach a FORM/SORM iterative technique, presented in chapter 2, is used
for the calculation of the probability of failure given by Eq.5.2. The safety margin (or
the limit-state function) can be expressed as

( \,,,
(5.5)
M - E - D L - I B,IB,;fE[SmJ
I. B)

where is the thickness correction exponent on stress, T is the thickness of the plate
and TB is the reference plate thickness.

In this approach z, K, , B f and Bgti can be modelled as random variables and any
appropriate probability distributions can be assigned to these variables. In a simplified
treatment of uncertainties in fatigue loading, all the uncertainties in environmental
description and hydrodynamic calculations are introduced through a single random
variable B511.. When all the loading variables have to be individually considered they
can be introduced implicitly in the calculation of the safety margin by making
& E[S] a function of these random variables. In this case the entire stochastic
response calculations have to be included within the iterative reliability calculations.
Alternatively, if these loading uncertainties can be expressed through the parameters
of the long-term Weibull model the safety margin expression can be modified and
written as
5-32

MA - D L E - I I + (1 -W) c tm ri( ^n11 (5.6)


K TB ) [ B) D)j

In this case the five Weibull model parameters W, A, B, C, and D can be treated as
correlated random variables.

Example 1: Fatigue Loading Uncertainties Expressed Through a Single Variable.

The two reliability analysis approaches presented above are demonstrated here with
an example. The tubular joint considered here is a part of an 6-legged jacket structure
in 60m water depth the details of which are given in Appendix-D. The long-term
wave climate is represented by 6 sea-states and two directions of wave approach are
considered.

Reliability analysis is first carried out using the lognormal formulation considering
i, K, Bg , and as random variables. The example joint has a mean fatigue life
of 540 years neglecting the effect of thickness correction and fatigue endurance limit.
The results of the analysis along with relevant input data are shown in Table 5.3 for
a 20 year service life (reference period).

The results show a relatively low reliability index for the joint considered which is
also indicated by a low mean fatigue life and happens to be one of the 10
fatigue-critical joints identified for the structure. The parametric sensitivity factors for
the various basic variables show that the reliability index is higly sensitive to the
variability in the model uncertainty parameter for the calculation of global transfer
function followed closely by the model uncertainty in the calculation of stress
concentration factors and the variability in S-N data. The uncertainty in Miner's
summation is of less importance to the computed probability of failure.
5-33
Table 5.3: Results of reliability analysis for the limit-state of fatigue
using an S-N approach and lognormal formulation.

Reliability index 3 = 3.138


Probability of failure Pr = 8.5E-04
Joint No: (Elem. 3346 / Node. 303)

Variable Distribution Mean COV atvav

__________ Iognormal 1.00 0.30 -0.509


K lognormal 1.012E+13 0.58 -0.803
Bgtr lognormal 1.00 0.30 -4.585
B3 r lognormal 1.00 0.25 -3.920

Reference period = 20 years


Plate thickness = 80 mm
Slope of S-N curve = 3.0
Cumulative Loading = 3.748E+l 1

The high contribution of the variability in B51. and B 1 to the total variability in
fatigue life and the high sensitivity of 3 to these variables should be evident from the
exponent m2 for these variables appearing in the calculation of 0lnM in Eq.5.3. The
high sensitivity to the uncertainty in B g u should be of concern because veiy little
information is available about this variable to develop accurate statistics and it is
generally very difficult to reduce the variability in B 51 . The use of lognormal model
for this variable is also largely hypothetical. The sensitivity to these assumptions can
be investigated using the generalized formulation.

First, the reliability analysis for the joint is repeated using the first order reliability
method using the same probabilistic description of variables as before and the results
are presented in Table 5.4. The same value of reliability index of 3.138 is obtained.
Since all variables are lognormal the failure surface will be linear in the standard
normal space and the FORM analysis gives the exact result in this case. The direction
cosine sensitivity factors again confirm the high sensitivity of 1 3 to
5-34
Table 5.4: Sensitivity of f3 to the distribution assumption for Bgd•


Reference period = 20 years
Reliability index = 3.138 (3.301)
Probability of failure = 8.5E-04 (4.6E-04)

Variable Distribution Mean COV a

lognormal 1.0 0.30 0.225


__________ _______________ ___________ ___________ (0.260)
K lognormal 1.012E+13 0.58 0.413
___________ ________________ ____________ ____________ (0.477)
Bgff lognormal 1.00 0.30 -0.676
___________ (normal) ____________ ____________ (-0.525)
B f lognormal 1.00 0.25 -0.567
(-0.655)

Next, in order to test the sensitivity of the results to the lognormal assumption for Bgd
the analysis was repeated using a normal distribution for this variable but with the
same level of variability and the results are given in braces in Table 5.4. The results
indicate a small increase in the reliability index but a considerable change in the
sensitivity factors. The model uncertainty in the stress concentration factor now
becomes the most dominant variable. It is however possible to reduce the variability
in B f considerably by using even a 2-D finite element analysis. The reliability
analysis was repeated again using a smaller COV of 0.25 for Bg (which can be
obtained by improving the stochastic analysis procedure and a more accurate
environmental description) and a COV of 0.15 for B5 (a value typical of a 2-D finite
element analysis). This gives a reliability index of 3.764 showing the significant
benefit that can be obtained by reducing the variability of the most important
variables. With the reduced variability the variable B is seen to be of less
importance than the variable K which has a relatively high variability. Finally,
repeating the base case analysis given in Table 5.4 but also including the variability
in thickness correction gives a reliability index of 1.864 for the joint.
5-35
Example 2: Loading Uncertainties Expressed Through the Weibull Parameters.

A more detailed reliability analysis is now carried out using a bi-linear S-N curve and
the loading uncertainties are modelled through the 5 parameters of the long-term
Weibull model. The correlations used for the Weibull parameters are: PAB=PcD=-0.6'
PAC=PBD=°•4' and PAD=PCB=-°•3 while W was modelled as independent of others.
In addition, the thickness correction exponent and the stress level S0 at which a
change in slope of the S-N curve occurs are modelled as random. The slope in 2 is
taken as 5 and K2 is calculated as K2=K1S0m2l. The results are given in Table 5.5.

A deterministic assessment of the joint including thickness correction gives a mean


fatigue life of 270 years and a nominal life of 80 years based on the D.En. 'T' curve.
The reliability analysis, relating again to a service life of 20 years, incorporates a more
realistic representation of uncertainties in fatigue assessments and the reliability index
of 1.847 is typical of joints with estimated lives of 80 years based on a design S-N
curve. A study of the sensitivity factors again shows that the uncertainty in fatigue
loading and calculation of stress concentration factors are the dominant variables. The
high sensitivity of reliability index to uncertainties in Weibull parameters A and B in
comparison with C and D derives from the fact that the weighting factor is 0.63 and
therefore the first part of the mixed-Weibull density function contributes most to the
stress-range density function. However, when the stress spectra are bi-modal, as in
most deep-water jacket platforms the contribution of the second part of the density
function will be significant and in such cases the uncertainty in parameters C and D
will also be important. The reliability index is seen to be less sensitive to the
uncertainty in total number of stress cycles as the variability is generally low.
Similarly, J3 is relatively less sensitive to the uncertainty in thickness correction and
stress level S0 in comparison with other resistance variables.

5-36

Table 5.5: Results of reliability analysis for the limit-state of fatigue


using an S-N approach and generalized formulation.

Reliability index 3 = 1.847


Probability of failure Pr = 3.24E-02
Joint No. (3346/303)

Variable Distribution Mean COV a

lognormal 1.00 0.30 0.228

K lognormal 1.012E+13 0.58 0.366

normal 0.30 0.15 -0.147

S0 normal 90.0 0.30 -0.001

lognormal 1.0 0.25 -0.502

A lognormal 1.0* 0.20 -0.644

B lognormal 1.0* 0.15 0.354

C lognormal 1.0* 0.20 -0.042

D lognormal 1.0* 0.15 0.009

W lognormal 0.63 0.15 -0.074

lognormal 1.0 0.10 -0.068

* bias (Units in N, mm)

The reliability analysis methodology presented above can also be used to derive safety
factors to be used with determinitic fatigue life estimates. A reliability analysis for
the joint considering o1y the uncertainties in S-N curve and thickness correction and
a reference period of 80 years, which is the same as the estimated life, gives a
reliability index of 1.82. This means that the D.En. T curve, which is 2 standard
deviations below the mean S-N line gives an implicit reliability index of 1.82 which

5-37
is reasonably adequate to cover the uncertainties in S-N data. However, when all the
loading and other uncertainties, typically involved in a fatigue analysis, are considered.
as can be seen from the above example a safety factor of approximately 4 on life is
required to maintain the same level of reliability as implicit in the 'T' curve.

A plot of reliability index versus the factor to be applied on estimated fatigue lives
is given in Fig.5.l. From this plot safety factors for any target reliability can be
obtained. Thus, considering a target reliability of 2.0 (p 1.=0.0227) for an inspectable
joint the required safety factor is 5 which means that for a platform service life of 20
years the estimated fatigue life of the joint should be greater than 100 years.

Tar&et Reliability

2.5

1.5

0.5

0 1 I I

0 2 4 0 8 10 12 14
Factor on Fatigue Life

Fig. 5.1: Variation of safety factor on fatigue life with target reliability.

Similarly, choosing a target reliability of 3.00 (p= .O.00l) for an inspectable underwater
joint of major importance the required factor on life is 25 and for a reliability of 3.70
(jO.00Ol) for an un-inspectable joint of major importance the factor is 70. These
factors seem to be considerably higher than the commonly used factors of between
3-10. However, when the uncertainties in response analysis can be reduced by using
more refined methods and using finite element methods for SCFs the required factors
on life can be reduced.
5-38
5.6.2 Reliability Analysis using a Fracture Mechanics Approach:

In the development of an S-N curve the criteria commonly used to define failure by
fatigue is the formation of a through-thickness crack. Consequently, reliability
analysis can be carried out for this failure event. However, in a fracture mechanics
approach, since the fatigue crack growth under service loading is explicitly calculated,
reliability analysis can be carried out for a limiting crack depth less than the plate
thickness. Any value of allowable crack depth can be specified from considerations
of accessability for inspection or repairability of the cracked joint. Thus a safety
margin equation for the limit-state of fatigue can be formulated as, see also Shetty and
Baker (1990c)

M - a-a(r) (5.7)

where a is the limiting crack depth (for example plate thickness) and a(t) is the crack
depth after a service exposure of time t. Starting from an initial crack depth of a0 the
crack depth a(t) after time t can be calculated using the fracture mechanics crack
propagation model presented in section 4.4.4. In this case a(t) is a function of the
random variables such as initial defect size, fatigue material properties, uncertainties
in service loading, etc. Alternatively, using the expressions for the fatigue resistance
function (Eq.4.40) and the fatigue loading function (Eq.4.41) developed earlier, the
safety margin for fatigue failure can also be expressed as

M - lvR(aC)-WL(t)

(5.8)
- ItBRIIBSIE[Sm]
CGB5YU[7rx/q] m 2 -

where G, "'a and Ua are all functions of the instantaneous crack depth and are already
defined in chapter 4.

In the above formulation for safety margin the parameters a 0, C, B511, B 1, and B 11 can
be modelled as random variables to account for the uncertainties in initial defect size,
fatigue material properties, stress intensity factor model, stress concentration factors
5-39
and the response analysis under fatigue loading, respectively. In addition the
uncertainty in threshold stress intensity factor range IKth can be introduced through
G and the uncertainty in residual stresses through U. The safety margin is thus
implicitly defined in terms of the basic variables and a simple lognormal format
cannot be used for reliability analysis of this limit-state. However, the FORM
iterative method can be used efficiently as demonstrated in the following example.

Example 3: Variation of Reliability With Service Exposure.

The tubular joint considered in example 1 is re-analyzed here using a fracture


mechanics approach. Maintaining the same service load history and using the crack
propagation model developed in chapter 4 gives a mean fatigue life of 287 years
which is very close to that obtained using the S-N approach. Properties have been
slightly adjusted for this joint to obtain nearly equal mean lives from the two
approaches. A reliability analysis is carried out considering the plate thickness as the
limiting crack depth and the results are given in Table 5.6. All the variables in this
example are considered as statistically independent of each other except for the
Weibull parameters A and B which are assigned a correlation of PAB=-°•6• The other
parameters C,D and W are not treated as random variables as they are not important
for the joint considered, see Table 5.5.

The reliability index for the joint is obtained as 1.747 which is significantly lower than
that obtained using the S-N approach, although the mean lives obtained by the two
approaches are nearly equal. This can be attributed to the additional uncertainty
involved in the calculation of degree-of-bending factors and stress intensity factors
required in a fracture mechanics approach and the fact that the reliability index is
highly sensitive to these variables. If the variability in Bb and B 1 is ignored (or
reduced to zero!) the reliability index can be increased to 2.536. This value is slightly
higher than that obtained using the S-N approach as the variability in material
parameter C is relatively smaller than that of the S-N parameter K, other variabilities
being the same.
5-40

From the direction cosine sensitivity factors given in Table 5.6 it can be seen that the
reliability index is highly sensitive to the model uncertainty in the response
calculations, SCFs, SIIFs and DoBs and relatively less sensitive to the variability in
material parameter C. The uncertainty in initial defect size, threshold factor and
residual stresses do not have a significant influence on the probability of joint failure
by fatigue. It is interesting to note that the variable BdOb acts as a "resisting" variable.
This is because increase in BdOb results in an increase in the bending component of the
through-thickness stress thus reducing the total stress intensity factor and thereby
increasing the fatigue life.

Table 5.6: Results of the reliability analysis for the limit-state of fatigue using a
fracture mechanics approach.

Reliability index = 1.747


Probability of failure = 4.03E-02

Variable Distribution Mean COV a

Weib-A lognormal 1.0 0.20 -0.640

Weib-B lognormal 1.0 0.15 +0.364

B 1 lognormal 1.00 0.25 -0.474

B 5f lognormal 1.00 0.20 -0.379

BdOb lognormal 1.00 0.15 0.298

Paris-C lognormal l.OE-1 1 0.40 -0.247

a0 lognormal 0.15 0.66 -0.088

(a/c)0 lognormal 0.62 0.66 -0.006

normal 1.60 0.20 -0.00t

S lognormal 300 0.20 -0.000


5-41
Table 5.7: Improvement in reliability with reduced variability.

Reliability index = 1.954 (2.128) [2.318]

Variable COV a2 X 100

Bgtt• 0.30 (0.30) [0.25] 55 (64) [55]

B 0.15 (0.15) [0.15] 14 (17) [20]

B511 0.20 (0.15) [0.15] 18 (12) [15]

BdOb 0.10 (0.00) [0.00] 6 (0) [0]

C 0.40 (0.40) [0.40] 7 (7) [10]

The statistics used in Table 5.6 corresponds to a simplified design situation where
parametric equations are used for the calculation of stress concentration factors, degree
of bending factors and stress intensity factors. Next, the improvement in reliability
that can be obtained by successively refining the analysis techniques and thereby
reducing the variability in model uncertainty parameters is studied. First, the
variability in and BdOb is reduced to correspond to a situation where the stress
analysis of the intact joint is carried out using finite element methods while the stress
intensity factors are calculated using parametric equations. This increases the
reliability index to 1.954 and the order of importance of variables changes as shown
in Table 5.7. In the next step the variability in B 5 is also reduced to correspond to
a situation where the stress intensity factors are calculated using finite element
methods, such as line springs or 3-D f.e.m in which case degree of bending factors are
not used. This gives a further increase in j3 to 2.128 and the sensitivity factors for
this case are shown in braces in Table 5.7. Finally, assuming that a more accurate
environmental description is used and a rigorous stochastic response analysis is carried
Out, the variability in B5 is slightly reduced. This, however, gives a significant
improvement in reliability ([=2.3l8) as B51. happens to be the most important
variable. The results for this case are enclosed in square brackets in Table 5.7.
5-42

In conclusion, it can be said that when a fracture mechanics analysis is used for
fatigue assessment it is important to have a reasonably good estimate of the stress
distribution of the intact joint. This can be obtained by a marginal increase in
computational effort but it will give a significant increase in the computed reliability.
The uncertainties in the calculation of response transfer functions to determine
nominal member stresses are of paramount importance in fatigue assessment and effort
should be made to reduce these uncertainties by improving the quality of analysis.

In addition to the value of the reliability index for a joint, which can be used to assess
the adequacy of the design, what is of more interest is to generate information about
the variation of reliability through the service life of the structure. This can be used
for judiciously planning the inspection times for various joints.

Reliability Index
4,

3.5

- - -
2.5

1.5

0.5

0

0 2 4 6 B 10 12 14 16 18 20
Service Exposure (years)

Fig.S.2: Variation of reliability index with time for the fatigue limit-state
5-43

The variation of reliability index with service exposure for the joint considered earlier,
using a fracture mechanics approach, is shown in Fig.5.2. The reliability index is as
high as 3.50 at the end of 2 years of service exposure but drops to a value of 1.747
at the end of 20 years, considered to be the design life of the structure. On this plot
possible target level for inspection is indicated, which more appropriately should be
recommended by regulatory authorities. From this plot the time for first inspection
can be obtained as 8 years. After the inspection, and depending on whether a crack
is detected and sized and accounting for the reliability of the inspection techniques,
the reliability of the joint can be updated using Bayesian updating methods.
Depending on the updated reliability and its post-inspection variation in the remaining
service life, a decision on the time-to-next inspection can be made. Bayesian updating
techniques are, however, not considered in this thesis.

5.7 RELIABILiTY ANALYSIS FOR LIMIT-STATE OF PLASTIC COLLAPSE

In a deterministic approach for the design against plastic collapse, for example based
on Dept. of Energy guidance notes (1990), the joint capacities are estimated using
characteristic equations derived from test results on scaled specimens. A joint is
deemed "safe" if the maximum forces caused by a 100-year return wave condition do
not exceed the characteristic strength. No additional safety factor is used to guard
against the event of environmental conditions exceeding the 100-year values and the
capacity of the joint becoming less than the characteristic value. This has been
justified in the background document to the guidance notes, see Dept. of Energy
(1989b, OT.H308), on the consideration that the response calculations and design of
tubular members is based on the working stress approach while the joint capacity
equations have been derived from ultimate strength tests. Reduction in joint capacity
due to possible fatigue crack growth is not accounted as the fatigue strength is
separately checked for. However, in view of the considerable uncertainties involved
in the prediction of design environmental conditions and considering the empirical
nature of strength equations a reliability based approach for design against plastic
collapse of tubular joints is desirable and this is presented in the following.
5-44

When a tubular joint is subjected to the combined action of axial forces and bending
moments through the brace, based on Eq.4.7, a limit-state equation for joint collapse
can be formulated as

( BenvPa " (Beny Ma ' (Be,wMa'l


M - 1 - (5.9)
B01P B 01 M B01M U
J+L j7PB

in which, Pa (M)p and (Ma)OpB are the applied force and moments on the joint and
P (M) and (MU)OPB are the joint capacities calculated using the mean-value
strength equations presented in chapter 2. The variables B 1 and are introduced
here to account for the model uncertainties in the joint strength equations and in the
calculation of wave kinematics and hydrodynamic loading on the structure. The
collapse loads P (and similarly (MU)JPB and (MU)OPB) and the applied forces P (and
similarly (M) J1B and (Ma)OPB) can be expressed as functions of additional random
variables as

- PU(FY,aO,cO,C,m,.Klh.......)
(5.10)
- Pa(H,T,dw,Vc,tmgCa,Cm,...)

in which, in addition to the yield stress, the joint capacities are expressed as functions
of random fatigue parameters to account for the reduction in joint capacity due to
random fatigue crack growth over a service exposure of duration t. Similarly, the
applied forces are expressed as functions of random environmental parameters such
as wave height, period, water depth, current, Morison coefficients etc. Using the
response models of chapter 3 and strength models of chapter 4, reliability analysis for
the limit-state of joint collapse can be carried out using the FORM method of
chapter 2. This is demonstrated below with an example.

Example 4: Reliability Analysis of an Intact Joint.

First, reliability analysis is carried out for a joint under life-time extreme
environmental condition without including any fatigue deterioration. The joint has a
utilization factor of 0.7 under mean-value environmental loading. During iterative

5-45
search for the beta-point the random environmental loading is determined using a
response surface approach which is discussed in more detail in the next chapter. The
wave height and period are considered as correlated with a correlation coefficient of
Pm =°•4 while all other variables are considered as statistically independent. The
results of the reliability analysis are presented in Table 5.8.

Table 5.8: Reliability analysis for the limit-state of plastic collapse

Reliability index = 5.003


Probability of failure = 2.86E-07
Joint NO. (3346/303)

Variable Distribution Mean COY a

H Gumbel 18.0 m 0.15 -0.692

T lognormal 11.0 sec 0.10 -0.278

tmg normal 50.0 mm 0.15 -0.25 1

Cd lognormal 0.7 0.20 -0.323

Cm lognormal 1.8 0.20 -0.001

lognormal 1.0 rn/sec 0.20 -0.021

lognormal 1.0 0.15 -0.249

F lognormal 300 MPa 0.10 0.175

B lognormal 1.00 0.15 0.428


5-46

The reliability index of the joint is 5.00 which is relatively high in comparison with
the fatigue limit-state. From the direction cosine sensitivity factors it can be seen that,
of the environmental parameters, the uncertainty in wave height and drag coefficient
have the greatest influence on the reliability index. Almost negligible sensitivity to
the inertia coefficient implies that the extreme wave loading for this particular
structure is drag dominated. This may not be so for a different structure or under
fatigue condition when the inertial loading is expected to be significant. The model
uncertainty variable Benv which accounts for the uncertainty in hydrodynamic force
calculations, although has a significant influence on reliability, it is relatively less
important than in the case of fatigue (compare with the sensitivity factor for B g in
Table 5.6). The reliability index is also highly sensitive to the uncertainty in the
calculation of joint capacities. Therefore, higher research priority should be given to
improve the present understanding of collapse behaviour of tubular joints and develop
more reliable strength prediction equations.

Next, the analysis is repeated for the joint allowing for fatigue deterioration over a
duration of 20 years and including all the fatigue random variables in the reliability
analysis. However, for this joint inclusion of fatigue did not cause any significant
influence on joint reliability and the sensitivity factors for the fatigue variables were
found to be close to zero. The reliability index for the limit-state of fatigue failure
for this joint is 1.75. Thus for this joint, although the dominant failure mode is
fatigue, the reliability of the joint against plastic collapse is not influenced by fatigue
deterioration. This pattern was observed for most other joints in the structure
considered but the results cannot be generalised to all cases. A different conclusion
may be obtained for a joint with very low fatigue life and for a wave direction for
which the applied forces due to extreme loading are also high.
5-47
5.8 RELIABILITY ANALYSIS FOR THE LIMIT-STATE OF FRACTURE

The current industry practice relies entirely on the choice of appropriate material
properties for the parent plate and welding consumables for fracture avoidance in
offshore structures. Explicit fracture assessments are only made when defects larger
than those prescribed by quality control standards are discovered during fabrication
or in-service inspections. However, fracture failure is a complex process involving
the interplay of a number of factors and it is desirable that fracture failure is treated
as an explicit limit-state during the design process. In view of the considerable
sources of uncertainty involved in the fracture assessment of tubular joints a reliability
based approach is more appropriate. In the following, two approaches are presented
for reliability analysis for the limit-state of fracture corresponding to the two
commonly used fracture assessment models.

5.8.1 Reliability Analysis Based on CEGB R6 Fracture Assessment Model:

In section 4.5.2 of chapter 4 a deterministic fracture assessment model has been


proposed for tubular joints which is based on the C.E.G.B R6, Milne et al (1985),
failure assessment diagram approach. This model can be extended to probabilistic
fracture assessment by treating the major sources of uncertainty through a set of
random variables and determining the probability of joint failure by fracture. The
failure assessment curve, defined by Eq.4.64 and shown in Fig.4.18, can be considered
to constitute a failure surface in the KrLr space, corresponding to which a safety
margin expression can be formulated as

M-K -K T
(5.11)
- [l-O.l4L] [O.3+O.7exp(-O.65Lf)] - K,
5-48

where L1 is a plastic collapse parameter, K is a fracture parameter and K, is the limit


value of K 1 to cause fracture failure. The parameters and L 1 can be expressed as
deterministic functions of other basic random variables as

Kr - .11 [a(t),c(t),Sm,Syes,Km,Bsjj,...]
(5.12)
L - f2[a(t),c(r),Sm,Fy,Bcoj,...J

where, the functions f1 and f2 can be obtained from the relations presented in
section 4.5.2. In addition to the variables previously defined the variables B 1 and B1
are introduced here to account for the model uncertainties in the calculation of stress
intensity factors and joint collapse loads, respectively.

In the above, the crack dimensions a(t) and c(t) at the time of the extreme load
application could be taken as the random initial weld defect dimensions a 0 and c0
when fatigue is not considered or, more appropriately, be taken as the dimensions of
a fatigue crack after a service exposure of duration t. In this case the crack depth aCt),
and similarly c(t), can be expressed as functions of random variables affecting fatigue
crack growth as

aCt) - f3[aO,cO,C,AKSh,B8f,BSCf,BdOb,...J (5.13)

The functional relation f3 is given by the crack propagation model developed in


section 4.4.4 and the additional model uncertainty variables B 5Lf, B Cf, B dOb and B If are
introduced for reasons explained in section 5.6.2.

The maximum hot-spot stress Sm which induces fracture failure in a joint is


considered to be caused by the action of an extreme wave and can be expressed as a
function of the random environmental parameters and model uncertainty parameters
as

S,,,, - f4[H,T,dw,Vc,tmgCd,Cm,BenvBzcfBdob••.] (5.14)


5-49
The function f4 can be obtained from the hydrodynamic loading and response models
presented in chapter 3, or more efficiently, using a response surface approach
presented in chapter 6. Here it is important to distinguish between the two loading
regimes, namely: (i) the fatigue loading, representing the cumulative action of all
waves over a duration t, which causes a weld defect of initial size a0 to grow by
fatigue to a size a(t) and (ii) the extreme loading which is considered to act at the end
of the reference time t and induce the hot-spot stress S which causes failure of the
joint by fracture. The fatigue stresses are usually determined using a spectral response
analysis while the response under an extreme wave event is determined using a
discrete-wave static analysis. Some of the basic variables may be common to the two
loading regimes and the final choice of basic variables to be included in a reliability
analysis and their probabilistic modelling depends on how the "extreme load event"
is defined and the method of response analysis used for the fatigue condition and the
extreme condition.

As an alternative to the above formulation of the extreme load event, it is possible to


represent the extreme stress Sm by the highest stress peak that could occur in the
fatigue stress spectrum. As a simplification it may be assumed that the highest stress
peak occurs in the most severe storm sea-state out of all the sea-states used to
represent the long-term wave climate. Of the n individual stress peaks arising in a
sea-state lasting for a duration t', and when stress peaks are assumed to follow the
density function given by Eq.3.25, it can be shown that the maximum stress peak S
follows an extreme-value Type-I distribution. The mean value and standard deviation
of can be obtained as, see Thoft-Christensen and Baker (1982)

- [(21n(v 0 t')]' 12 +
(5.15)
- it/[121n(v0t')]'12
5-50
where y=O.5T12 (Euler's constant) and Z=S m /(Y in which c is the r.m.s. value of the
stress process. This type of reliability formulation and an example based on this is
given in Shetty and Baker (1990c). In this case a separate response analysis for the
extreme load event is not required and the loading random variables will be common
to both for the determination of cyclic stress-range S for fatigue and the peak stress
S for fracture.

A further important consideration in the above analysis is the time t corresponding to


the extreme load event at which fracture failure occurs. In the above, the reference
time t is specified by the analyst and it is assumed that the extreme load event will
occur at any time within the ItthII year but the fatigue crack growth is conservatively
calculated to the end of the tth year. Thus the environmental parameters in Eq.5.14
should be the annual extreme values and the Sm in Eq.5.15 should be the highest
stress peak caused by the annual extreme sea-state. Using this approach reliability
analysis can be carried out for various exposure periods and the variation of annual
probability of failure through the service life of the structure can be studied. This type
of analysis, although it does not give the life-time probability of failure, has a
considerable utility in a design assessment to ensure that the annual reliability level
of the joint does not deteriorate (due to fatigue) below acceptable limits. Similarly,
decision to inspect or repair may be made based on the annual probability of failure.

One disadvantage of the above approach, however, is that the effect of fatigue
deterioration on the reliability of a joint under a life-time extreme environmental
condition, which in principle can occur at any time in the service life of a structure,
cannot be assessed. What is of considerable interest is to determine the most likely
combination of fatigue deterioration and extreme loading which will cause failure of
the joint by fatigue-fracture and the probability of such an event occurring within the
service life of the structure. Such an alternative reliability formulation is presented
in chapter 7 where time-dependant issues are more rigorously addressed.
5-51
Example 5: Reliability Analysis for the Limit-State of Fracture:

Reliability analysis is first carried out for the limit-state of fracture without including
any fatigue deterioration. Fracture failure of a tubular joint is strongly influenced by
the level of tensile residual stresses. The reliability analysis is therefore carried Out
for a joint in both the as-welded condition and the stress-relieved condition. The
results of reliability analysis for a joint in the as-welded condition, assuming mean
residual stresses of 300 MPa, are shown in Table 5.9. Next, the analysis is repeated
for the stress-relieved condition, assuming mean residual stresses of 30 MPa, and the
results for this case are shown in braces in Table 5.9.

From the sensitivity factors for the as-welded condition it can be seen that the failure
behaviour is dominated by residual stresses rather than the applied primary stresses
but the reliability index is relatively high compared to other limit-states for the joint.
The environmental parameters which contribute mainly to the primary stress do not
show any influence on the joint reliability while the sensitivity of f3 to residual
stresses, fracture toughness and the model uncertainty in the calculation of stress
intensity factors is very high. Compared to fatigue limit-state it can be seen that the
uncertainty in initial defect size is relatively more important for the fracture limit-state.
The reliability index is less sensitive to B C as it contributes to primary stresses alone
while the joint failure in this case is governed by secondary stresses.

However, in the stress relieved condition the fracture behaviour is dominated by


primary stresses associated with high plasticity. Consequently, the environmental
parameters and loading uncertainties have a strong influence on joint reliability. The
reliability index is 4.92 which is very close to the value of 5.00 for the limit-state of
plastic collapse. The sensitivities are also similar to that of the plastic collapse mode.

5-52

Table 5.9: Results of reliability analysis for the limit-state of fracture.

Reliability index = 4.342 (4.928)


Probability of failure = 7.07E-06 (4.OE-07)
Joint No. (3346/303)

Variable Distribution Mean COV a

H Gumbel 18.0 m 0.15 -0.147 -0.610

T lognormal 11.0 sec 0.10 -0.154 -0.365

tmg normal 50.0 mm 0.15 -0.042 -0.154

Cd lognormal 0.7 0.20 -0.104 -0.394

Cm lognormal 1.8 0.20 -0.000 -0.000

V lognormal 0.9 rn/sec 0.20 -0.020 -0.059

lognormal 1.0 0.15 -0.064 -0.242

a0 lognormal 1.0 mm 0.50 -0.187 -0.002

S lognormal 300,3OMPa 0.25 -0.434 -0.042

K lognormal 8OMPaIm 0.30 0.711 0.118

B f lognormal 1.0 0.25 -0.056 -0.047

B 1f lognormal 1.0 0.20 -0.480 -0.079

B dOb lognormal 1.0 0.15 0.011 0.003

F lognormal 300 MPa 0.10 0.032 0.226

B lognormal 1.0 0.15 0.061 0.423


5-53
Next, the reliability analysis is repeated allowing fatigue deterioration in the joint over
increasing exposure periods. In this case, the random crack size a(t), after fatigue
crack growth over a specified duration 1, is calculated as in Eq.5.13 using the fatigue
crack propagation model. The results of reliability analysis for various exposure
periods are shown in Table 5.10. The sensitivity factors are shown only for important
variables.

Table 5.10: Reliability analysis for fracture failure with various durations of
fatigue exposure (Joint No. 3346/303).

t 0.0 2.0 4.0 8.0 12.0 16.0 20.0


J
13 4.342 3.960 3.642 3.282 3.162 2.954j 2.708

Variable a

H -0.147 -0.126 -0.121 -0.115 -0.109 -0.104 -0.097

WA -0.000 -0.018 -0.023 -0.042 -0.044 -0.047 -0.049

C -0.000 -0.065 -0.081 -0.109 -0.112 -0.123 -0.134

a0 -0.187 -0.457 -0.432 -0.402 -0.398 -0.379 -0.362

B f -0.056 -0.166 -0.210 -0.254 -0.265 -0.273 -0.288

B 11 -0.480 -0.499 -0.524 -0.546 -0.554 -0.562 -0.574

S -0.435 -0.365 -0.356 -0.342 -0.337 -0.332 -0.328

K 0.711 0.592 0.578 0.560 0.558 0.552 0.548

The reliability index decreases continuously as the fatigue exposure increases. From
the variation of sensitivity factors it can be seen that the influence of environmental
parameters remains low and decreases gradually with increase in exposure period.
The sensitivity of the reliability index for variables which are common to both fatigue
5-54

and fracture phenomenon such as B1 and increase with fatigue exposure. The
sensitivity to fatigue variables such as WA and C increases while that for fracture
variables such as residual stresses and fracture toughness decreases with increase in
exposure period. From these results it is clear that with the increase in exposure
period the role of fatigue variables becomes more important for the joint reliability.
Similar conclusions were obtained in Shetty and Baker (1990c).

6 Reliability Index

2
Fatigue Failure
-e- Fatigue—Fracture
1
- Plastic Collapse
0
0 2 4 6 8 10 12 14 16 18 20
Service Exposure (years)

Fig.5.3: Variation of reliability index with time for all limit-states

Finally, the variation of reliability index with service exposure for the limit-states of
fatigue, plastic collapse and fatigue-fracture is shown in Fig.5.3. From this it can be
seen that, for the joint considered, the reliability is dominated by fatigue limit-state
throughout the service life of the structure. The variation of reliability under
fatigue-fracture limit state is more gradual and always remains higher than the fatigue
limit-state of through-thickness cracking.
5-55
5.8.2 Reliability Analysis Based on BS PD 6493 Fracture Assessment Model:

The reliability analysis formulation presented above can also be applied when the
BS PD 6493 fracture assessment model, see BSI (1989), is used. This should be the
preferred route when the material toughness data is available in terms of crack tip
opening displacement. The latest draft revision of BS PD 6493 using a CTOD
concept is in fact very similar to the CEGB R6 approach except for a difference in the
failure assessment curve as discussed in section 4.5.1.5. Based on Eq.4.60 a safety
margin expression for reliability analysis using the BS PD 6493 model can be
formulated as

Mj-fj
1/2 (5.16)
- s 1__lnsecui.S
2 r)1
where S7 is a plastic collapse parameter, 6. is a fracture parameter and 3.. is the limit
value of 8 for fracture failure. The parameters S. and 6 can be obtained from
Eqs.4.61 and 4.62 given in section 4.5.1.5. In order to account for the various
uncertainties in loading and material parameters in a reliability analysis the parameters
S and 3 can be expressed as functions of other basic random variables in a manner
similar to Eqs.5.12-5.15, discussed earlier in relation to the CEGB R6 model. The
probability of joint failure by fracture can be calculated using the FORM method as
demonstrated with an example below.

Example 6: Reliability Analysis Based on BS PD 6493 Model for Fracture.

The reliability analysis of the joint considered in the previous example is now carried
out using the BS PD 6493 approach. The value of the material CTOD is obtained
from the K, value using the conversion formula 6=(K 2)/(F E) to enable a
consistent comparison between the two approaches. The results are given in Table
5.11, where the sensitivity factors are shown for important variables only.
5-56

Table 5.11: Reliability analysis for the fracture limit-state using BS PD 6493.

Reliability index = 4.4 12


Probability of failure = 5.13E-06
Joint No. (3346/303)

Variable Distribution Mean COV a

H Gumbel 18.Om 0.15 -0.142

T lognormal 11.0 sec 0.10 -0.134

Cd lognormal 0.7 0.20 -0.108

a0 lognormal 1.0 mm 0.50 -0.119

S lognormal 300 MPa 0.25 -0.454

lognormal 0.20 mm 0.30 0.724

B$Cf lognormal 1.0 0.25 -0.056

B51f lognormal 1.0 0.20 -0.468

F lognormal 300 MPa 0.10 0.084

B lognormal 1.0 0.15 0.064

The reliability index is slightly higher than that obtained earlier using the CEGB R6
approach. The sensitivity factors are nearly the same in both cases.
5-57

5.9 CONCLUDING REMARKS

The treatment of uncertainties for reliability analysis of tubular joints should consider
the uncertainties involved in modelling both the fatigue loading and the extreme load
event. The models used for stress analysis and strength assessment of tubular joints
are subject to considerable uncertainty and often these model uncertainties dominate
the computed reliability of a joint, and their accurate modelling is vital to reliability
analysis. More data collection and rigorous statistical analysis for fatigue material
properties, fracture toughness, welding defects and residual stresses cannot be
overemphasised, as the reliability of tubular joints is governed more by the parameters
which affect the resistance of a joint than by the uncertainties in the prediction of
environmental conditions. The final choice of basic variables to be included in a
reliability analysis depends on the type of response analysis method used for
determining fatigue and extreme response, on how the joint limit-states are formulated
and the level of computational simplification desired.

Although fatigue failure is the dominant failure mode for a number of joints in an
offshore structure, it has little effect on the reliability of many joints for which the
dominant failure mode is either plastic collapse or fracture under an extreme wave
condition. For some joints which are critical in both fatigue and fracture (or plastic
collapse) the combined effects are important and the reliability of such joints is
sensitive to both fatigue and fracture variables. For the selection of joints to be
included in a system reliability analysis it is useful to categorise the joints as:
(i) fatigue-critical, (ii) plastic collapse-critical, (iii) fracture-critical and (iv) combined
mode and assign corresponding failure modes to these joints. This will reduce the
computational costs considerably and avoid convergence problems during FORM
iterative analysis. However, it is to be noted that the above categorisation will be
influenced by such factors as: (a) the assumed direction of approach of the extreme
wave in relation to the dominant direction of fatigue waves, (b) the reference time of
analysis or the service exposure duration and (c) whether a single value of fracture
toughness is used for the HAZ and parent plate or, if different values are used, the
extent of FIAZ and the relative toughness values of the HAZ and parent plate.
5-58
In a system reliability method which involves the analysis of a large number of failure
sequences every effort should be made to keep the number of basic variables to a
minimum. The total computational time increases drastically with the total number
of basic variables included in the analysis. From the results of reliability analyses for
isolated tubular joints carried out in this chapter, the following variables are identified
to be the most important for the three modes of joint failure:

Fatigue: Crack growth parameter C, initial defect depth a 0, model


uncertainty in SCF calculation B$Cf, model uncertainty in DoB
calculation BdOb, model uncertainty in SIF calculation B1 and
the uncertainty in the calculation of stress response transfer
function Bgff.

Plastic Collapse: Yield stress F. model uncertainty in the calculation of collapse


loads B 1 and the environmental parameters for the extreme
load event such as wave height H, period T, drag coefficient Cd,
inertia coefficient Cm, thickness of marine growth tjflg. and the
model uncertainty in the calculation of hydrodynamic loading
and response

Fracture: Material fracture toughness Kmat, residual stresses S, the


initial defect size a0 , and the model uncertainty parameters B,
B dOb, BSIf, and B 1 mentioned above.

Finally, careful consideration should be given to how the individual joint limit-states
are formulated, the definition of the extreme load event and the reference time of
analysis. These have a significant bearing on how the results of the reliability analysis
are interpreted and used as decision tools in design assessments or in-service
inspection planning of jacket-type offshore structures.
Chapter 6
SYSTEM RELIABILITY ANALYSIS UNDER AN EXTREME WAVE

6.1 GENERAL

In chapter 5, methods for reliability analysis of individual tubular joints were


presented without consideration of the collapse behaviour of the remainder of the
structure. Similarly, formulations can be developed for the reliability analysis of
tubular beam-column members for the limit-state of plasticity or buckling failure. For
a statically determinate structure, failure of any one of its members or joints will cause
structural failure. In such cases, the probability of failure of the structural system can
be defined by the union of individual failure events of all members and joints in the
structure. The probability of the union event can be determined from the probabilities
of the elemental failure events, duly accounting for the correlation between failure
events, using the methods for ideal series systems discussed in chapter 2 (section
2.6.2). This system reliability will be less than (or equal to) the reliability of the most
critical component in the structure.

In the conventional deterministic design practice consideration is usually limited to the


limit-state of initial failure. Structural safety is ensured by designing individual
elements to meet certain minimum levels of safety. In most structures, such as
offshore jacket structures, additional redundancy is built in to avoid catastrophic
failure of the structure by the failure of any single element. However, mere
introduction of more redundant members and measuring redundancy by the number
of extra degrees-of-freedom (statical indeterminacy) may not always lead to high
reliability. If the uncertainty in loading is high, such as in extreme environmental
loading, compared to resistance uncertainty, the elemental failure modes will be highly
correlated and presence of redundant members may not give additional benefits. On
the other hand, when the resistance uncertainty is large, such as in the case of fatigue
failure of tubular joints, the reliability with respect to system failure will be much
higher than that of individual element reliability. Thus, in order to properly quantify
6-2
the degree of redundancy and the likely safety benefits arising from providing more
redundant members, system reliability methods have to be used.

In a redundant structure several elements have to fail before structural collapse occurs.
After the failure of the first element, i.e. when it reaches a limit-state, the load shared
by this member depends on its post-limit capacity. Thus, in a system analysis
post-limit behaviour is very important and this is discussed for tubular members and
joints in section 6.2. After the initial failure, the structure will continue to support
increasing load due to existence of alternate load paths and load-effect re-distribution
from failed elements to surviving elements until the structure is no longer able to
support any additional load and fails by the formation of a mechanism. Some of the
methods for progressive collapse analysis of structures are reviewed in section 6.3.
For efficiency of computation in a reliability analysis the use of a response surface
approach for determining linear response to environmental loading is discussed in
section 6.4. A new reliability formulation for analysis of a single failure sequence
involving only member failures under a life-time extreme wave event is presented in
section 6.5. This formulation involves the determination of the joint r3-point for the
failure sequence directly using a multi-constraint optimisation technique while the
progressive collapse behaviour of the structure is traced very efficiently using a virtual
distortion method. This methodology was developed within the BRITE project and
towards which, in addition to the author, a few other researchers have contributed.
A redundant structure will have a large number of possible failure paths but in general
only a small number of failure sequences contribute significantly to system failure
probability. The available methods for identification of probabilistically dominant
failure sequences are reviewed and a new failure-tree enumeration method for practical
analysis of large structures is developed in section 6.6. Finally, in section 6.7, the
proposed methodology is extended to include fatigue deterioration and failure of
tubular joints by plastic collapse or fracture. The various formulations presented in
this chapter have been implemented by the author in association with other researchers
in the reliability analysis software package RASOS developed under the CEC BRITE
Project. The proposed methodology and the RASOS software is demonstrated in this
chapter on a full-scale jacket structure.
6-3
6.2 LIMiT AND POST-LIMIT BEHAVIOUR OF MEMBERS AND JOINTS

The post-limit behaviour of members and joints is an aspect of considerable


importance in system reliability analysis. The post-failure strength of individual
elements not only influences the ultimate strength of the structure but also the type of
structural collapse analysis method that can be used and to some extent the system
reliability formulation, as will be clear from the remaining sections in this chapter.
Some of the commonly used post-limit capacity models for tubular members and
joints are discussed below.

6.2.1 Models for Tubular Members:

Various idealised models have been used for the limit and post-limit behaviour of
structural members in the system reliability studies reported in the literature. The
advantages and disadvantages of some of these models are discussed in the following.

(a) Elastic-Ideal Plastic, Semi-Brittle Truss Member Model:


The simplest approach to modelling member behaviour is to use a bi-linear
load-deformation behaviour as shown by curve 1 in Fig.6.1. In this the behaviour is
assumed to be linear-elastic until the peak capacity, P, is reached and thereafter the
member is assumed to retain this load level but undergoes increasing deformation.
If unloading occurs at any stage, the unloading stiffness is assumed to be the same as
the initial stiffness in the linear-elastic range. The peak capacity in tension can be
taken as the plastic capacity of the section while the peak capacity in compression can
be determined from Euler buckling theory or using any code formulae depending on
the member slenderness. This type of member model has been widely used in early
system reliability studies, see for example Murotsu et al (1980).
6-4

LOAD

DEFORMATION

Fig.6.1: Idealized post-limit behaviour for a truss element

However, after a buckling failure in compression the load-carrying capacity of a


member generally reduces. This can be modelled by an ideal-brittle behaviour as
shown by curve 3 , see Murotsu et al (1980), or more generally by a semi-brittle
behaviour as shown by curve Z in Fig.6.1. In this, after the failure of a member, its
load carrying capacity is assumed to drop instantaneously from its peak capacity P
to a residual capacity TP without any increase in deformation. Note that T=1
corresponds to ideal-plastic case and Tl=O to ideal-brittle case while in most cases the
post-limit compression capacity is modelled by assuming r to be between 0.3-0.8.
This type of modelling has been used by Moses (1982), Guenard (1984) and
Nordal et al (1987).

The main advantage of this model is that it is simple to use in a structural collapse
analysis using successive elastic analysis, see section 6.3. Since the post-limit
behaviour is independent of deformation it is not necessary to monitor deformations
or strains in failed elements. However, in reality most members strain-harden under
tension while the capacity reduction after buckling failure is more gradual with
increasing deflection. Suck a muhi-state truss model has recently-been used in a
reliability analysis in Rashedi and Moses (1988).
6-5

(b) Elastic-Ideal Plastic Frame Member Model:


A frame model in which the members are represented as beam-columns gives a more
realistic representation of most structures. In conventional design of frames the
material stress-strain behaviour is often represented as elastic-ideal plastic. This gives
a bi-linear moment-rotation characteristics for members as shown in Fig.6.2a. The
axial load-deformation may also be assumed to be bi-linear for simplicity. Such ideal
ductile member behaviour allows the use of plasticity theorems for collapse analysis
of frames and has been used by Watwood (1979), for deterministic limit analysis of
frames, Ditlevsen and Bjerager (1984) and also Thoft-Christensen and Murotsu (1986),
for system reliability analysis.

AXIAL
MOMENT LOAD

ROTATION MOMENT
(a) (b)

Fig.6.2: Idealized post-limit behaviour for a frame element

When a member is subjected to the combined action of axial load and bending
moments it is necessary to consider interaction effects. The influence of combined
loading on member capacity is usually represented in terms of an interaction curve as
shown in Fig.6.2b. The interaction curve may be defined by the von-Mises or Tresca
yield criteria or in practice by interaction equations given by codes, eg. API (1987),
which are developed from experiments. It is common to assume that yielding or
plasticity occurs in a member at a discrete number of sections, such as at the mid-span
6-6
and ends or under concentrated loads. The post-limit behaviour of a yielded section
('plastic hinge') is defined by the 'flow rule' or the 'normality condition' which
states that for a given stress-state on the limit-surface the vector of plastic strain
increments has the direction of the outward normal to the limit-surface. Thus under
increasing deformations the relative magnitudes of axial load and moments change
such that the coordinate point (P,M) moves along the limit surface.

Rigorous analysis of this flow rule based frame model is quite difficult and requires
an incremental or iterative procedure at the plastic hinges. In a reliability analysis the
use of such incremental methods is very expensive and simplifications are often made.
In Murotsu et al (1981) the effect of axial loads is ignored, so that effectively there
is no interaction, while in Badshauge and Bach-Gansmo (1985) the interaction effect
is considered but the flow rule is ignored. In Thoft-Christensen and Murotsu (1986)
a linear limit surface is used which considerably simplifies the introduction of flow
rule. In Turner and Baker (1987) a non-linear interaction with the associated flow rule
is used for the reliability analysis of a small structure. The effect of neglecting the
flow rule is in general to under-estimate the ultimate load of the structure and hence
to overestimate the probability of failure, Turner and Baker (1987). The use of a
linear limit surface, similarly, under-estimates the peak capacity of the member and
the collapse load for the structure.

(c) Generaliseci 3-D Beam-Column Member Model:

The use of an elastic-ideal plastic member behaviour, although it offers considerable


simplicity in structural analysis, misrepresents the true behaviour. Most members
strain-harden under tensile loading while they loose their load-carrying capacity after
a buckling failure under compression. A more accurate approach is to use a
generalised beam-column model as shown in Fig.6.3.- Under axial tension the member
is assumed to strain-harden as represented by the hardening angle a in Fig.6 3a, while
under axial compression the member after reaching the peak capacity is assumed to
soften as given by the softening angle cz.
6-7

AXIAL
MOMENT LOAD

ROTATION MOMENT
(b) (c)

Fig.6.3: Generalized beam-column model for a tubular member

The effect of in-plane and out-of-plane moments is considered while the contribution
of shear forces and twisting moments, which are usually small, is neglected. The
moment-rotation characteristics is assumed to be ideal-ductile as shown in Fig.6.3b.
The interaction between axial load and moments is represented by a non-linear limit
surface as shown in Fig.6.3c. In general the behaviour should account for flow rule.

The above member model gives a more accurate representation of behaviour of tubular
braces and leg members of a jacket structure. The limiting values of loads and
6-8

moments for tubular members can be calculated using code formulations, such as
given in API (1987), with the safety factors set to unity. The interaction curve for
combined loads is also given by the codes. These formulations are developed from
test results on a large number of tubulars of varying slenderness and implicitly account
for the effect of residual stresses, initial imperfections etc. However, the hardening
and softening slopes are usually not specified in the codes and these have to be
assumed based on judgement or test results. Within the CEC BRITE Project a more
rigorous non-linear member model has been developed which calculates the
compressive load-carrying capacity and the softening slope a using an elasto-plastic
buckling analysis. The model also explicitly accounts for initial out-of-straightness
of a member and the occurrence of plasticity can be investigated at specified discrete
number of stations along the member. The member is assumed to be supported on
end rotational springs which model the effect of stiffness of the rest of the structure
and the stiffness of the tubular joints at the ends. The detailed theory and validation
has been reported in Abassian et al (1989) and is not described here.

Since the post-limit behaviour of the member is strain dependent, the above model can
only be used within a general purpose non-linear collapse analysis computer program.
A progressive collapse analysis program RASOS_C based on a virtual distortion
method has been developed within the BRITE project, see next section for details.

6.2.2 Models for Tubular Joints:

The limit and post-limit behaviour of tubular joints depends on the failure mode being
reached, i.e. plastic collapse or fracture. For the plastic coflapse limit-state the
load-deformation characteristics have already been shown in chapter 4, Figs.4.2 and
4.3. From this it can be seen that, under axial loading, after initial linear behaviour
the true load-deformation characteristics becomes non-iinear before the ultimate load
is reached. After the ultimate load is attained this load level is maintained while the
joint undergoes plastic deformations, but at large strain levels the load-carrying
capacity begins to drop. While the loss of capacity under axial compression is due
6-9
to chord wall buckling, under axial tension the joint loses its strength due to the
formation of a crack at the hot-spot and an eventual separation of the brace from the
chord. Under in-plane and out-of-plane bending moments the overall behaviour is
similar to that under axial loads. The load-deformation characteristics are influenced
significantly by joint geometry and configuration.

AXIAL LOAD L LOAD

MOMENT


DEFORMATION MOMENT
(a) (b)

Fig.6.4: Idealized post-limit behaviour for tubular joints

At the present state of tubular joint technology, formulations are not available to
describe the entire load-deformation characteristics. Empirical equations are presently
available to determine joint stiffness (flexibility) only within the linear-elastic regime
and these can be used to describe end restraints for tubular braces for linear elastic
structural analysis. However, the non-linear stiffness characteristics of tubular joints
can have significant influence on member behaviour and more research is required to
quantify this effect. For the time being the load-deformation and moment-rotation
characteristics can be approximated by a bi-linear behaviour as shown in Fig.6.4a.
The limit loads and moments can be determined using the formulations presented in
chapter 4 (section 4.2). The axial load-moment interaction curve is shown in Fig.6.4b.
and is defined by Eq.4.7. For the fracture failure mode the behaviour is idealised by
a semi-brittle model as shown in Fig.6.4a. The limit values of axial loads and
moments are governed by the crack size, fracture toughness and stress state at fracture
and can be determined by the fracture assessment model proposed in chapter 4.
6-10
6.3 STRUCTURAL COLLAPSE ANALYSIS

In a redundant structure the failure of a single element does not lead to the failure of
the structure. After the failure of the first element, the load shared by this member
is limited by its post-limit behaviour while the surplus forces are re-distributed to the
remaining elements and the structure continues to support increasing loads until a
sufficient number of elemental failures occur and the structure fails by the formation
of a mechanism. In recent years there has been a growing interest in the study of
collapse behaviour of jacket structures to estimate the ultimate capacity and the true
level of redundancy or residual strength after the failure of one or two members. Such
information will be of vital importance in the design of new type of structural
configurations and in the assessment of existing older platforms. In the system
reliability analysis of jacket structures an efficient tool for collapse analysis is vital.
The available methods for structural collapse analysis can be grouped into three
categories, namely: (1) direct plastic mechanism analysis, (2) successive elastic
analyses and (3) non-linear progressive collapse analysis. These methods are reviewed
in this section mainly from the stand point of deterministic collapse analysis. The use
of these methods in a system reliability analysis requires further modifications and this
is considered in the next section.

6.3.1 Direct Plastic Mechanism Approach:

For elastic-ideal plastic member behaviour, limit analysis can be carried out for truss
or frame structures using the upper bound (kinematic) or lower bound (static)
theorems of plasticity, see for example Heyman (1971). The upper bound theorem is
more commonly used and in this method it is necessary to identify a number of
possible plastic mechanisms for the structure using an appropriate method. Provided
all possible mechanisms itre identified, the true mechanism and the ultimate load for
the structure corresponds to the mechanism with the lowest load factor.
6-11
For a structure all possible plastic mechanisms can be written as linear combination
of a set of independent mechanisms called 'fundamental mechanisms'. If the
structure is r-degrees statically indeterminate and has n locations of potential plastic
hinges, each with k degrees-of-freedom, then the number of fundamental mechanisms
is given by m=n k-r. The fundamental mechanisms are entirely determined by the
geometry of the structure and the location of potential plastic hinges. For simple
structures these mechanisms can be identified by inspection but for more complex
structures these can be generated by the method given by Watwood (1979) which is
amenable for computerisation.

Once a set of fundamental mechanisms have been identified, the next step is to
combine these to form new plastic mechanisms for the structure. This process can be
carried out manually or more efficiently by linear programming methods. For each
mechanism, using the principle of virtual work, relations can be written between the
axial forces and bending moments at the hinges and the virtual displacements and
rotations. Using the equilibrium conditions for the structure the internal forces and
moments at the hinges can be expressed in terms of the applied external loads. The
solution of these equations gives the collapse load for the particular mechanism. The
procedure is repeated for all possible mechanisms and the lowest load is taken as the
true collapse load for the structure.

The advantage of the method is that it is theoretically rigorous given the assumption
of ideal plastic member behaviour. However, more realistic member behaviour such
as tensile hardening/compression softening or non-linear behaviour of joints and
fracture failures cannot be treated using this approach. Moreover, partial failure of
only one or two members leading to intolerable deflections which may be of more
interest to practising engineers cannot be assessed using this method.
6-12
6.3.2 Successive Elastic Analyses Approach:

This is one of the widely used methods of progressive collapse analysis within system
reliability studies. The steps involved in the analysis can be briefly described as
follows. First the intact structure is analyzed for the applied loading using the
methods of linear-elastic structural analysis. If there is no failure of any of the
structural elements the loading is increased to cause the failure of the most critical
element and the corresponding load factor is determined. At this stage the failed
element is "removed" from the structure, i.e. its stiffness is set to zero. The effect of
the member is accounted for by applying artificial loads corresponding to the
post-limit load of the failed member. Thus, for a truss element, forces equal in
magnitude to the residual capacity of the failed element are applied at the end nodes
of the member and these are kept fixed (but may be of random magnitude in a
reliability analysis) during the analysis for subsequent failures. If the initial forces in
the member are tensile the artificial forces will also be of same sign, see for example
Moses (1982) or Thoft-Christensen and Murotsu (1986). Similarly, for a frame
element, when a plastic hinge develops at one end of the member, stiffness terms
corresponding to the degrees of freedom at this end are set to zero and artificial forces
and moments, equal in magnitude to the post-failure capacity of the section, are
applied on either sides of the failed section. Because of this the method is often
referred to as a 'member replacement' method.

In the next step, since the remaining structure is still elastic, linear structural analysis
can still be used. Treating the applied loads and the artificial forces as separate load
cases, a reanalysis is carried out to determine new forces in all the surviving elements
and the load is incremented to cause the failure of the second member. The process
is repeated, each time replacing a failed element with artificial forces, until the
structural collapse occurs after the failure of Q elements.

Let R=(R, R2 ...R1 J be the vector of strengths of the members failing in succession
and the corresponding vector of load increments be r=(r1 , r2,...,r,). Then, the system
behaviour can be descnbed by, see Moses (1982) and Meichers and Tang (1984)
6-13

(6.1)
GR -Ar

where A is the 'utilisation matrix" and G is the 'load shedding' matrix. The matrix
A represents the relationship between the member strengths and load increments. For
the first element to fail only the first load increment is involved; for the second node,
only r1 and r2 are relevant, etc so that A is lower triangular. Similarly, G is a lower
triangular matrix with elements G representing the increase in force in member j due
to a unit reduction in post-limit capacity (for brittle and semi-brittle elements) of the
failed member i. For ductile structures there is no reduction in capacity, so G=I.
Inversion of expression (6.1) yields
(6.2)
r- A'GR
and the collapse load for the system is given by

R3 - r, (6.3a)
i-I

for ductile structures and

R - max[r1,r1+r2r1+r2+r3...] (6.3b)

for brittle structures.

This method, because of its simplicity and since computer programs for elastic
analysis are readily available, has found increased use in system reliability studies but
it suffers from following inherent limitations.

(a) Since a failed element is "removed" from the structure and its post-failure
deformations are not monitored, the method can be used only for members
with strain-independent post-limit behaviour. Thus for example ideal plastic
or semi-brittle member models discussed in section 6.2 can be used while more
complex models with strain hardening/softening behaviour cannot be used.
6-14

(b) The method assumes proportionally increasing loading, i.e. a set of nominal
loads are multiplied by a single load factor which is monotonically increased.

Cc) Since the deformations in failed elements are not monitored unloading of a
failed element following brittle failure of some other element cannot be
determined. This may lead to incorrect estimation of the ultimate load and the
sequence of failures may also be misrepresented, see Karamchandani (1990).

(d) Complex failure scenarios such as plastic failure of tubular joint at a member
end followed by buckling failure of the brace member cannot be investigated.

6.3.3 Non-Linear Progressive Collapse Analysis Approach:

In principle, progressive collapse analysis of an offshore structure can be carried out


using any of the conventional non-linear finite element program. However, the use
of these programs can be extremely expensive as each member and joint may have to
be modelled using a large number of finite elements. For example, Ueda and
Rashed (1991) report the use of 8 finite elements, each with 40 integration points
along the circumference, to model one member in one of their comparison studies.
For this reason, in recent years a number of computer programs have been developed
for efficient collapse analysis of framed structures in which each structural member
or joint can be modelled by a single finite element. Some of the programs used in
offshore applications include: USFOS program developed at NTH, Norway, see
Soreide et al (1986); INTRA program from PMB systems, USA, see Lloyd and
Clawson (1984); SAFJAC program from Billington Osborne-Moss Engineering, UK,
see Lalani and Shuttleworth (1990) and NOAMAS program from Osaka university,
Japan, see Ueda and Rashed (1991).

The basis of all the above programs is same. They all use incremental solution
techniques of one form or the other. For example, in USFOS program the
Euler-Cauchy incrementation method is used in combination with iterative equilibrium
6-15
corrections at each load level. For this purpose, the elasto-plastic response of each
structural element is first developed either from beam-column theory, empirical
solutions or from experimental data. The member response is usually expressed in the
form of load-deformation (and moment-rotation) curves and axial load-moment
interaction curves. Using these an incremental stiffness matrix is developed for each
element from which the global stiffness matrix for the structure is generated.

The loading is applied incrementally and each element is then checked for buckling
and/or plastification. Since the stiffness of the element is dependent on deformation
and internal forces, a new stiffness matrix is constructed and transformed into global
coordinates for each element after each load increment. The global stiffness matrix
is reassembled and the next increment of load applied, Ueda et al (1991). After
plastification or buckling failure of any element the internal forces are controlled to
lie on the limit surface. The collapse of the structure is detected by excessive plastic
deformation. From this approach a complete sequence of elemental failures and a
load-deformation curve for the structure can be developed from which the ultimate
capacity of the structure can be determined.

The predictions from these programs have been compared with those obtained from
non-linear finite element programs and good agreement is seen. These methods
generally give considerable computational savings over non-linear finite element
programs. However, the process of formulating the incremental stiffness matrix for
each element and for the structure at each load step and the solution of non-linear
simultaneous equations can still be very expensive for a large structure. Recently,
Holnicki-Szulc and Gierlinski (1989) have proposed a more efficient and general
approach for simulation of structural modifications, such as material redistribution or
changes to constitutive relations of elements. The approach is called 'Virtual
Distortion Method' (VDM) and it forms a vital element of the system reliability
method proposed in section 6.5. A broad outline of the VDM approach is given
below for a truss structure and closely follows that of Holnicki-Szulc and
Gierlinski (1989).
6-16
The virtual distortion method is based on the concept of superposition of 'original
state' of the structure with a 'virtual state' to obtain the 'modified state' of the
structure. In the context of progressive collapse analysis, the original state of the
structure refers to the linear-elastic, stress-strain state, which is known, and the
modified state refers to a stress-strain state of the structure with one more failed
elements, which is to be determined. The "failure" of elements is simulated by
introducing a set of virtual distortions into the failed elements and determining the
virtual stress-strain state of the structure. Then the stress-strain state of the modified
structure can be expressed as

CM - CL + Lv
(6.5)
- L +

where C L, L - linear strains and stresses in the original structure,


C V, - virtual strains and virtual stresses caused by virtual distortions.

Strains CV and stresses a, are defined as follows:

-
(6.6)
- E(D-J)

where D is the influence matrix with element D describing the deformation of the
member i caused by a unit distortion =1 of the member j,
I is the unit matrix, and
E is a vector of elastic material properties.

Substituting virtual strains and stresses from Eq.6.6 into Eq.6.5 gives


CM CL + DE
(6.7)
aM cYL + E(D-J)
6-17
Once a component reaches a limit-state its stress is constrained by its post-limit
behaviour. If the stresses from a linear-elastic solution exceed the post-limit stress,
the surplus stress correction can be expressed as a set of virtual distortions at the
failed members. Based on Eq.6.6, the virtual distortions can be calculated as

Bi -

- -(a-a) (6.8)
B LTE(DI)

where =sign(a) is the vector of post-limit stresses and a is the vector of current
total stresses in failed elements and thus Aa gives the surplus from the member. L
is a 'participation matrix' with non-zero elements corresponding to all failed
members and thus B, called the 'active influence matrix' is a sub-matrix of D
corresponding to failed members only. The above equation is formulated assuming
that the total virtual distortions and total stresses a are updated each time a new
member fails. The solution of Eq.6.8 gives the virtual distortions , which, when
substituted in Eq.6.7 gives the required stress-strain state of the modified structure.

If the load is applied in an incremental manner, strains, stresses and virtual distortions
have to be accumulated at the end of each load increment. If more than one elements
fail during a load-step, additional procedures have to be adopted to detect the sequence
of failures based on the stress gradient at the previous load level and the
corresponding algorithm of computational steps is given in Holnicki-Szulc and
Gierlinski (1989).

The considerable computational efficiency of the VDM approach derives from the fact
that the set of simultaneous equations which have to be solved to obtain the vector of
virtual distortions (Eq.6.8) is usually small (about 10-15) compared to the size of the
system as virtual distortions are introduced in only failed elements. The linear-elastic
stiffness matrix and the linear solution which is readily available from any of the
standard linear analysis programs can be used. Compared to the incremental
approaches, which require formulation of incremental stiffness matrix and solution of
the system of equations for the entire structure at every load step, the VDM method
6-18

gives enormous computational savings. The only limitation is that since the method
is not incremental, in principle, the elasto-plastic response of an element should be
given in a piece-wise linear form. However, any general non-linear response
characteristics (which in any case are rarely known exactly) can be accurately
specified in a piece-wise linear form.

The above methodology has been generalised within the CEC BRITE project to frame
behaviour with softening/hardening effects and tubular member and joint elements.
The methodology has been implemented into the collapse analysis program
RASOS_C, see Gierlinski and Lau (1991), which is one of the modules of the RASOS
system for reliability analysis of offshore structures.
6-19
6.4 RESPONSE SURFACE MODEL FOR ENVIRONMENTAL LOADING

Based on the member and joint models and the virtual distortion method for
progressive collapse analysis a reliability analysis methodology is developed in the
remainder of the chapter. In addition some of the commonly used reliability analysis
procedures are reviewed and their inherent limitations pointed out. But first a
discussion on the determination of random linear response of an offshore structure is
necessary and this is considered next.

The loading on a jacket structure consists of gravity loads (self weight, appurtenances
and deck loads) and those imposed by the environmental action of wind, wave and
current. The uncertainty in gravity loads, which is usually small, can be modelled
through a random multiplier applied directly to the response (e.g. nodal forces)
because of the linear relation between loading and response. The modelling of
uncertainties in extreme environmental loading has already been discussed in
section 5.2. In general, the internal nodal force (TN ) and moment (MN) vectors can
be expressed as a function of the uncertainty variables as

TN , MN - f[H,T,d,VC,V,t,Cd,C,...] (6.9)
where, H is the wave height, T is wave period, d is water depth, V is current speed,
V is wind speed, tmg is thickness of marine growth, Cd is drag coefficient and Cm is
the inertia coefficient. The function f[.] comprises models for water particle
kinematics (for example, Airy or Stokes theory), calculation of wind, wave and current
loads using Morison approach and finally structural analysis to determine stress
resultants from nodal loads. Because of the complexity of these models it is usually
not feasible to include these calculations within a system reliability analysis using
simulation or iterative methods.

In a number of studies on system reliability of offshore structures reported in literature


the environmental loading is often highly simplified. For example, in Nordal
et al (1987) the environmental loads are expressed as equivalent nodal loads and this
load pattern is assumed to remain constant. The nodal loads are scaled by a single

6-20

load factor which is expressed as a simple function of wave height. The uncertainty
in the load factor is derived from the uncertainty of wave height. This method allows
the treatment of uncertainty in wave height alone and makes the nodal forces all over
the structure to be fully correlated. On the other hand, in Murotsu et al (1985) the
wave loading is expressed as point loads on members the statistics of which are
derived numerically from the variability in wave height, drag and inertia coefficients.
This approach will be computationally very expensive for a large structure, both for
the determination of load statistics and in reliability analysis as the number of basic
variables will increase considerably. Within the CEC BRITE project, Labeyrie et al
(1992) have developed a response surface approach for determination of response to
random environmental loading. The objective here was to develop an analytical
function for f[.] in Eq.6.9 above which is reasonably accurate and computationally less
expensive to use. This model has been implemented in RASOS software and it is
summarized below.

In the first step, the water particle kinematics are described using Airy's linear wave
theory as it provides a greater simplification in developing a response surface model.
Using a coordinate system with its origin at the sea surface and Z-axis increasing
upwards, the water particle velocity and acceleration in a direction normal to the wave
crest (x-direction) can be given as

H it
V(x,z,t) - - [cosh (kdz) cos (kdax-W)]
T sinh(kd)
(6.10)
H 2ic2
A(x,z,t) - - [cosh(kdz) sin (kdax-ii)]
T 2 sinh(kd)

where k, d, 1, a: wave number, water depth, width of structure, and a=l/d,


pseudo phase angle, and
z=(Z/d)+ 1.

The next important step consists of separating the particle velocity and accelerations
into a random component and a deterministic component. For extreme waves, the
variability in k is usually small and thus the hyperbolic functions in the above
6-21
equation can be developed using a first-order Taylor's expansion around the mean
value of wave number k, giving

H it
V(x,z) - [cos(w) V 1 (x,z) + sin(W) V2(xz)]
T sinh(kd)
(6.11)
H 2it2
A (x,z) - [cosCi) A 1 (x,z) + sin(W) A2(x,z)]
T 2 sinh(kd)

where, the vectors V1 , V2, A 1 and A2 are deterministic vectors defined as

V1 (x,z) - cosh(kdz) cos(kdax) U 1 + sinh(kdz) sin(kdax) u2

V2(x,z) - cosh(kdx) sin(kdax) u1 + sinh(/cdz) cos(kdax) U2


(6.12)
A 1 (x,z) - V2(x,z)

A2(x,z) --V1(x,z)

in which u 1 =OX cos(0) + OY sin(0) and u2=OZ where 0 is the wave direction.

The wave forces are computed in a normal way using the Morison equation. The drag
and inertia components of the force, per unit area, normal to a cylindrical member can
be given in local coordinate system (x',y',z') as

fD(x ) -

(6.13)
f1 (x) - PDCmA(X'L)

f (x') - fD(x') + f,(x')

where, p is the mass density of sea water and Dc=D+2tmg is the equivalent diameter
in which D is the member diameter and tmg is the thickness of marine growth. The
forces are integrated over the length L of the member.

The remaining steps in the development of the response function, although tedious,
are relatively straightforward. The wave force on each beam element is calculated by
integrating over the member length and transferring it as equivalent nodal loads at the
6-22

end of the member. The loads are then transformed to the global coordinate system.
The total nodal loads at each node is obtained by summing up the contributions of
each of the members framing into that node. The effect of current is accounted by
vectoria1Iadding the current velocities to the wave induced particle velocities in
Eq.6.12, see Labeyrie et al (1992). Similarly, the nodal loads due to the action of
wind on exposed members above the sea level are calculated separately. Finally, the
equivalent nodal loads, for a node N, can be expressed as

8
(6.14)
TN - ATNI

where, the 8 deterministic vectors T1, i=1,..,8 depend on the topology of the structure
in the neighbourhood of the node N and have the form, for a beam element NP

T3 - g 1j [L,D,t g ,c j , p j J g 2 [V,] i-1,...,8 (6.15)

The functions g 11, g21 and the Morison coefficients c1 depend on whether the vector T1
corresponds to drag or inertia component and whether perpendicular or parallel to the
member. The vectors V, for i=1,..,6 are obtained from the particle velocities and
accelerations of Eq.6. 12 after transformation to global coordinate system and similarly
for kinematics due to current and wind action, see Labeyrie et al (1992) for detailes.

Of the 8 deterministic vectors in the above, the first 4 (T1 : i=1,..,4) correspond to wave
to
loading (2 drag and 2 inertia components), T 5 and T6 to drag loading dueLcurrent and
T7 and T8 to wind drag loading. The corresponding random multipliers for i=1,..,8
are functions of random basic variables and are calculated as
6-23

H
x'-E_ T sinh(kdjC;tmg

H
T sinh(kd) ]2Citm:
x2-[

H1Cmtg
[T2 sinh(kd)
J
H
?.4 1Cm'tg (6.16)
2sinh(kd)]
[T

S
- I H
]2CdtmgVc
[i' sinh(kd)

H
6
T Sjflh(kd)]2 dmg

, 7 — cw v

).8 -C V

where V is the reference current speed, V is the reference wind speed, C and
Care the aerodynamic coefficients. The primes (') indicate components parallel to
a member. The marine growth factor is defined as OThg=tmg (D+2 tmg)/D. where,
'mg is the actual marine growth thickness for a member and is a marine growth
enhancement factor for the member diameter. 'tmg is considered as a random variable
which is identical all over the structure. Similarly, the hydrodynamic and
aerodynamic coefficients are expressed as c,=C1 c, where, C.L is modelled as a
random variable identical for all members in a structure while c can have different
values for each member. The deterministic vectors T 1 are calculated using the values
of tmg and c while the combined random loading is calculated based on 0mg and c.

Since the vectors T1 : i=1,..,8 depend only on the topology of the structure they need
to be calculated only once. These can then be transformed into 8 components of the
member internal forces F33: j=1..,2xndof, where ndof is the number of degrees of
6-24

freedom (6 for a frame element) at each node using matrix structural analysis.
Because of the linear relationship between T1s and F1 (assuming quasi-static response)
the random multipliers X, can also be applied directly to F 131. Thus, during iterative
reliability calculations the random linear response (stress resultants) to environmental
loading can be determined as

S1 - XF J j-1,...,ndof (6.17)

The random multipliers ? are calculated for each iteration using Eq.6.16 from the
current iterate of the basic random variables which is relatively inexpensive. Thus the
response surface approach provides a very efficient modelling of uncertainty in wave
height, wave period, water depth, wind speed, current speed, marine growth and
hydrodynamic and aerodynamic coefficients.

6.5 RELIABILITY ANALYSIS FOR A PRESCRIBED FAILURE MODE

An important task in the system reliability analysis of practical offshore structures is


the identification of, so called, 'dominant failure mechanisms'. Methods for this are
considered in section 6.7, where it will be seen that each possible failure mechanism
is first chosen, on some logical basis, and its probability of failure calculated before
its stochastic significance can be ascertained. Therefore, the problem of calculating
the probability of a prescribed failure mechanism is first studied in the present section.
The attention here is focused on failure mechanisms resulting from yielding/buckling
failures of primary members of an offshore structure under a life-time extreme wave
condition. Generalization of the above method to include tubular joint failures under
fatigue deterioration are considered in section 6.7.

The previously available methods for system reliability have used either a 'failure
mechanism approach' based on plastic analysis or a 'failure sequence approach'
based on successive elastic analysis of the structure. These methods are first reviewed
and their limitations explained. A new method, based on the 'joint beta point

6-25

cOncept' for reliability formulation and the virtual distortion method for non-linear
structural analysis, developed within the BRITE project is presented next and its
advantages highlighted.

6.5.1 Reliability Analysis Based on Plastic Mechanism Approach:

If the material and member behaviour is elastic-ideal plastic a system reliability


method based on plastic analysis discussed in section 6.3.1 can be used. This method
requires the generation of fundamental mechanisms for a structure which can be
performed by the method of Watwood (1979) discussed earlier. A real collapse
mechanism for the structure can then be formed by a linear combination of one or
more fundamental mechanisms. Reliability analysis for one such collapse mechanism
is discussed here.

Let the number of external loads Q acting on the structure be Q and let there be n
sections where potential yield hinges can occur. Let there be 3 stress resultants at
each yield hinge, axial force P and bending moments M and M 1. Then using the
principle of virtual work a safety margin equation for the i' fundamental mechanism
can be written as, see for example Thoft-Christensen and Murotsu (1986)

n 1
M -
j-i
.u.^M
P1 1
0 .+M
PYJ Yf PZj82) ) - E QrnS:m
rn-i
(6.18)

where, u, O are a set of compatible displacements and rotations at the yield hinges
and 6=f(u,83,,0)1 are virtual displacements of the external loads for the 1h

fundamental mechanism. These can be determined from the plastic analysis method
given by Watwood (1979). Some of the displacements and rotations above may be
zero as the summation is over all possible hinge locations. The section plastic
capacities N 3, and M. for the jtl plastic hinge can be calculated for a tubular
member as
6-26

-
(6.19)
M Mpzj -WF
pyJ pj

where F is the material yield stress, A 3 is the member cross-sectional area and W,,
is the plastic section modulus. The interaction between the axial capacity and the
plastic moments is usually given in the form of a yield condition and the plastic
deformations should be constrained to follow the flow-rule, Delmar et al (1988).

Next, a failure mechanism for the structure can be obtained by the linear combination
of two or more fundamental mechanisms. When k (<3 n-r) fundamental mechanisms
are combined the safety margin for the failure mechanism can be expressed as

k
(6.20)
M -
pJ J
u +M 0 +M
pyj yj
0 ) -
p21 2)
EQö
rn-I

where is a coefficient used to combine the i fundamental mechanism. If a


fictitious mechanism is formed the external work for this mechanism will be zero.

A reliability analysis for this failure mechanism can be carried out using the
FORM/SORM methods as for a simple elemental limit-state as explained in chapter 2.
In general the iterations are carried out in the U-space of independent, standard normal
variables and a transformation to the X-space of basic variables needs to be performed
before evaluating the safety margin.

The mechanism approach requires the reliability evaluation of a single limit-state and
the probability of the failure mechanism is obtained directly. This is unlike the failure
sequence based approaches, discussed next, which require reliability analysis for a
number of conditional events to determine the probability of a single failure sequence
and analysia of a number of failure sequences to estimatethe probability of a failure
mechanism. However, an important drawback of the failure mechanism approach is
that it can be used only for materials with elastic-ideal plastic properties. Mechanisms
involving buckling failures or tubular joint fractures cannot be analyzed.
6-27

6.5.2 Reliability Analysis Based on Member Replacement Approach:

This is one of the widely used methods of system reliability analysis at the present
time. Unlike the plastic mechanism approach, this method evaluates the probability
of a sequence of failure events leading to the collapse of the structure. In almost all
the reported applications successive elastic analyses with member replacement have
been used for progressive collapse analysis of the structure (see section 6.3.2).

Consider an ordered sequence (k: 1,2,3,...j,..,q) of elemental failures leading to the


collapse of the structure. Such a sequence has been variously called as 'failure
sequence', 'failure path' or 'failure mode'. Let EI1,2 J..1 denote the event that the
element j fails after the elements 1,2,..,j-1 have failed in that sequence. For system
failure defined by collapse of the structure, in the kth failure sequence, all the p
elements should fail. Thus the system failure can be expressed as an intersection of
componential failure events and the kth modal failure probability can be calculated as

P/k - P[E1 flE211 flE311,2 fl...flEqp i,2•..•q_ j ] (6.21)

In order to calculate this path failure probability it is necessary to calculate the


probabilities for the individual (conditional) events.

A safety margin expression for the failure of the first element can be written as

M1 -R1-S1
(6.22)
S1 - EaQ

where R1 is the resistance of element 1, S 1 is the stress resultant in element 1 and Q


is a set of Q external loads acting on the structure. The influence coefficients a 11 can
be determined by matrix methods of structural analysis. In general an element of a
frame structure will have more than one stress resultant and in such a case the failure
of the element (e.g. yielding) is expressed by a limit-state criterion (e.g. von-Mises
yield criterion). The probability of failure p=P[M1^O] can be calculated in the usual
way using FORM/SORM methods of chapter 2.
6-28
After the failure of the first element there will be a re-distribution of internal forces
among the remaining members of the structure. In order to compute this new
stress-state the failed member is removed from the structure (i.e. its stiffness is set to
zero) and its effect is accounted for by applying a set of artificial loads equal in
magnitude to the residual strength of the member and of the same sign as the previous
stress resultants in the member, see section 6.3.2. The structure is re-analyzed using
linear-elastic structural analysis, treating the external loads and artificial loads as
separate load cases, to compute the new stress-state. The failure of element 2 can
now be assessed based on the new stress resultants S 21 . The procedure is repeated
for other elements in the sequence, each time replacing a failed element by artificial
loads and reanalysing the structure to compute the updated stresses. Thus a safety
margin expression for the th element failure can be formulated as, see for example
Thoft-Christensen and Murotsu (1986)

1' (6.23)
M-R- I EaiYQ:J_ bT1 1 R 1 - brl 2 R2 - ... - bri1R_1
where the superscript (j) refers to the level of analysis after (j-l) failures and if
the post-limit strength reduction factors. The reliability index f3, the vector of
direction cosines a and the probability of failure p=P[M^O] can be calculated using
FORM/SORM where, in addition to the resistance variables of element j and external
loads the resistances of elements I ,2,..,j-1 are also to be treated as random variables.

The failure path probability p can be calculated from the methods of parallel systems
presented in chapter 2 (section 2.6.1, Eq.2.37). The correlations between the
individual safety margins can be computed from the vector of direction cosines a1.
Note that this analysis gives the probability of occurrence of one failure sequence
leading to the collapse of the structure. However, the same collapse mechanism can
be formed by a number of alternative sequences (in this case q!) involving the same
elements and the probability of the failure mechanism should in principle be calculated
as a union of all these failure sequence events. This should not be construed as a
major drawback of the failure sequence approach. The failure sequences involving the
same elements will be highly correlated and often only one or two sequences
6-29
contribute significantly to the failure mechanism probability and it is usually sufficient
to consider only these failure sequences.

The member replacement approach as presented above has a number of limitations as


discussed below. These limitations arise both from the nature of the member
replacement technique used for progressive collapse analysis and the way the
probability of the failure sequence is calculated using Eq.6.21.

(a) In the member replacement technique of progressive collapse analysis, since


the member stiffness matrix of a failed element is set to zero, the deformations
in these members cannot be monitored in the subsequent stages of the analysis.
Because of this, unloading of previously failed elements cannot be detected
and this can lead to an incorrect failure sequence or incorrect estimate of the
collapse load.

(b) Multiple failure modes for a single structural member developing within the
same failure sequence cannot be accounted. Thus, for example, the effect of
a tubular joint failure on the capacity of the brace member or the gradual
development of two or more plastic hinges in the same structural member,
resulting in complete loss of member post-limit capacity, cannot be studied.

(c) Since the deformations in failed elements are not monitored, sirain-dependant
post-limit behaviour models for members cannot be treated using the member
replacement technique. Thus, for example, elastic-ideal plastic, brittle or
semi-brittle member models discussed in section 6.2 can be studied while
members with general softening/hardening post-limit behaviour cannot be
studied. Such realistic member behaviours can only be treated with non-linear
progressive collapse analysis methods discussed in section 6.3.3.

(d) In the procedure presented above, in order to estimate the failure sequence
probability, reliability analysis is carried out separately for each conditional
failure events to identify their individual beta-points.
6-30

0 'N 1(u).O

Joint b.ta-polnt

- - -

UI

Fig.6.5: fliustration of physically inadmissible design point, Turner, et a! (1988)

In a FORM analysis this amounts to linearising the conditional failure surfaces


at their individual beta-points as illustrated in Fig.6.5 for a failure sequence
with two elements (1-42). Finally, when the probability of the intersection
event is calculated using Eq.2.37 the result corresponds to the area between the
approximating hyperplanes in the failure region. This approach is sometimes
called a 'crude FORM system reliability approach' and this will give
inaccurate failure probabilities if the conditional failure surfaces are highly
non-linear in the standard normal space.

(e) Although the crude FORM approach can be easily used with a member
replacement technique it poses considerable problems when using standard
non-linear collapse analysis methods. This is because in a member
replacement approach the conditional safety margin expression for the event
0fjth
member failure (Eq.6.23) is explicitly defined in terms of basic variables
and can be evaluated without the need to follow a-load path-which results-in
this failure sequence. Thus referring to Fig.6.5, the conditional failure space
M211 ^0 includes regions which cannot physically occur, Turner et a! (1988).
The event 211 is conditional on member 1 having failed. Thus by following
6-31

a deterministic load path the event 211 is physically undefined unless M1^O.
These events can, however, be analyzed using a member replacement technique
as the safety margin is explicitly defined.

When using standard non-linear collapse analysis methods it is necessary to


assume a load path, for example proportional loading. In such cases it is not
possible to locate physically inadmissible design points as calculations cannot
be undertaken in physically inadmissible regions. The individual design points
are only used as linearization points for mathematical convenience and are
eliminated when the final intersection is evaluated, Turner et al (1988). Note
that the failure region for the intersection event M^O is in the fully
admissible region and corresponds to basic variable realisations for which both
members 1 and 2 fail in that sequence.

6.5.3 Reliability Analysis Based on Joint Beta-Point Concept:

In order to remove the limitations of the member replacement technique and the crude
FORM approach, as discussed above, and to enable the use of sophisticated non-linear
progressive collapse analysis methods to treat more general and realistic member
behaviour a new method of system reliability analysis is presented in the following.
As mentioned earlier, this method has been developed within the BRITE project and
a number of researchers have contributed in this task. Individual contributions are
acknowledged explicitly wherever possible and the specific contributions of the author
of this thesis are highlighted.

Following the discussion in section 2.6.1.2 of chapter 2, it is clear that the accuracy
of the FORM estimate of the failure sequence probability p given by Eq.6.21 can be
significantly improved if the linearization of the conditional failure surfaces is carried
out at the so called joint beta-point, proposed by Hohenbichler ci al (1987), instead
of at their individual beta-points as in the crude FORM approach. The joint beta-point
is a good choice not only because the joint probability density function (u) for all
6-32
u on the failure surface is maximum at this point, but as shown by
Hohenbichier et a! (1987), approximation at the joint beta-point gives asymptotically
exact results. This approach, although in use for some time in the DnV's PROBAN
software, see for example Madsen (1987), has not been widely used for the analysis
of reliability against progressive collapse of structural systems. One complexity with
this approach is that an advanced multi-constrained non-linear optimization algorithm
is required for the search of joint beta-point and the conventional Rackwitz-Fiessler
type of iterative algorithms used for element reliability analysis cannot be used.

A search procedure for the joint beta-point of a failure sequence (1,2,..j,..,q) can be
formulated as a constrained optimization problem as

12
U S : minimize

(6.24)
with constraints:

j-1,2,...,q

where n is the total number of basic variables and T denotes the transformation of
variables from X to U-space and g(.) refers to the limit-state function for element j
conditional on the failure of 1,2,..,(j-l) elements in that sequence. AU constraints are
not necessarily active at the joint design point i.e. g(u 5)=O is not necessarily valid
for all j and it is sufficient to consider only (1^m^q) active constraints.

Any algorithm for non-linear constrained optimization can be used for the solution of
Eq.6.24, however, the NLPQL algorithm given by Schittkowski (1986) has been found
to very efficient and has been implemented in the RASOS software by Gollwitzer
(1991). The NLPQL algorithm is based on a sequential quadratic programming
method. At each iteration step, a sub-problem with a quadratic approximation of the
Lagrangian function and linearized constraints is formulated. The solutionot this
quadratic programming sub-problem is used to determine a new iterate and the process
continues until convergence is achieved.
6-33

The failure sequence probability as expressed by Eq.6.21 is calculated as

Pfk m(8;1?)


3: 5, - aTu for I (6.25)


R: p.1 - a1Ta for i,j-1,2,...,m

where the unit normal vectors a are calculated from the gradients of the individual
failure surfaces evaluated at the joint beta-point and 3 is the shortest distance from
the origin to the failure surface of element i linearized at the design point. The
rn-dimensional standard normal integral can be evaluated using the approximate
method given by 1-lohenbichier and Rackwitz (1983), see chapter 2 for details of the
computational procedure.

The constraint functions, i.e. the conditional safety margins, in Eq.6.24 can be
expressed in a general form as

(6.26)
M - RJ-SJ¼I,i2,...,IJ_J,H,T,VC,Cd,...) j-1,2,...,q

where is the resistance of element j and is the stress resultant in member j after
the failure of 1,2,...,(j-l) elements in that sequence. The stress resultant is now
implicitly defined in terms of the environmental loading variables and a
strain-dependant post-limit resistance of failed elements 1,2,...,(j-l). In certain cases
the member capacity (limit behaviour) may also be strain-dependant. The linear
response due to environmental loading can be determined using the response surface
approach discussed earlier while the nonlinear response can be determined efficiently
by the virtual distortion method of progressive collapse analysis. It is interesting to
note that the uncertainty in environmental loading parameters influences only the
linear response while the non-linear response is influenced by the uncertainty in
post-limit resistance of failed members. This fact can be advantageously utilized for
computational saving during iterative calculations, see Shetty et al (1991).

It is important to note that the joint beta-point is always in a physically admissible


domain and the multi-constraint optimization algorithm can be cleverly tailored to
6-34
generate successive iterates which are on the failure side of the individual failure
boundaries, see Fig.6.5. This numerical strategy has been developed and implemented
in RASOS by Gollwitzer (1990). Thus, a set of loading and resistance variables can
be found which will result in the failure of elements in the exact sequence prescribed
so that the virtual distortion method of non-linear analysis can be used. Although it
is possible to search for the joint beta-point for the complete failure sequence directly
(and thus requiring only one reliability analysis per failure sequence) experience
suggests that it is better to do the analysis in stages to avoid convergence problems.
Thus, referring to Eq.6.21, the beta-point u 1' and the probability of failure P[M1^0]
is determined frst and then the joint beta-point U 1,2' and the probability
PI(M1^0)r(M21] and so on. At each stage the results are stored and used as starting
solution to the next stage of analysis, see Gollwitzer (1991). This also provides
additional information on the probability of partial failure events and this is precisely
the way a failure-tree is developed in a search algorithm for dominant failure
sequences, see section 6.6.

The next important task involves the modification of the VDM (which is essentially
a detemiinistic collapse analysis method) for use in a probabilistic frame-work and the
coupling of the limit-state function for jacket analysis (calculation of linear response
to environmental loading, strength assessment of tubular members and joints,
progressive collapse analysis and calculation of safety margins) with the NLPQL
algorithm. This part of the work has mainly been carried out by the present author
and is discussed next.

6.5.4 Modified V.D.M Algorithm for Prescribed Failure Sequence Analysis:

A deterministic progressive collapse analysis using the virtual distortion method


(VDM) was presented earlier in section 6.3 3. This is based on the algorithm given
by Holnicki Szulc and Gierlinski (1989) for truss structures and later extended to
frame structures and implemented in the RASOS_C module by Gierlinski and
Lau (1991). However, this deterministic procedure needs to be modified for use in
6-35

a reliability analysis based on the failure-tree approach. The required modifications


are discussed below.

It is first necessary to understand the distinction between how the VDM algorithm
works and how a reliability analysis based on the joint beta-point concept should be
run. The comments below apply equally well to other deterministic non-linear
analysis methods.

(a) In a deterministic approach, the linear response to gravity and environmental


loading, based on mean or design values of parameters, is calculated first.
During subsequent non-linear analysis the spatial distribution of nodal forces
through out the structure is assumed to remain fixed while the forces are
scaled by a single load factor. The loading is increased in several increments
until structural collapse occurs. Thus it is usually not possible to know the
corresponding values of environmental parameters (wave height, period etc.)
which would cause collapse of the structure. However, in a reliability analysis
one needs to model the environmental parameters as random variables and thus
it is necessary to determine the values of these variables at the joint beta-point
which corresponds to structural collapse.

In the reliability analysis approach used in RASOS the environmental loading


is determined using the response surface approach discussed earlier in
section 6.4. The nodal loads and corresponding responses are first calculated
separately for the 8 deterministic components of environmental loading.
During iterative reliability calculations the 8 random multipliers A are
determined as functions of the current iterate of environmental parameters
using Eq.6.16 and the combined linear response is calculated as in Eq.6.17.
The NLPQL optimization algorithm iteratively determines the joint beta-point
values of the environmental parameters.

(b) A given set of values of loading parameters and resistance variables will result
in only one failure sequence and a deterministic VDM analysis will
6-36
automatically produce this failure sequence. Within a loading increment, an
iterative analysis is performed until all the surplus forces from failed elements
are re-distributed to the surviving elements. Then the next load increment is
applied and the process continues until collapse occurs. This procedure can
be considered to result in a 'physically admissible' failure mode. However,
in a reliability analysis the interest is always to evaluate the probability of a
pre-selected failure sequence. Furthermore, it is to be noted that when
estimating the probability of the failure sequence as per Eq.6.21 it is implicitly
assumed that all the remaining elements of the structure are safe. Thus the
prescribed failure sequence is defined conditional on the survival of the
remaining members in the structure.

In a reliability analysis for prescribed failure sequence (1,2,...,q) the VDM


algorithm is therefore modified to introduce failures (more appropriately
termed state-changes) in only those elements which are on the failure
sequence. For this purpose, first the strength of member 1 is calculated and
compared with the imposed forces in the member computed from linear-elastic
analysis. If the limit-state criterion is violated the internal forces are adjusted
to follow the post-limit behaviour and the surplus force from member I is
simulated by a set of virtual distortions and the updated stress-state in the
remaining members calculated. Next the failure of element 2 is checked and
if it fails the surplus force from elements 1 and 2 are re-distributed and so on
until state-changes are introduced in all the components of the failure sequence
precisely in that order. If for a particular iteration the internal stress-state does
not result in the failure of one or more components of the prescribed failure
sequence the iteration is discarded and the NLPQL algorithm generates a next
iterate. This iterative process continues until the joint beta-point is found
which results in the desired failure sequence. Thus the possible failure of
other elements which -are not on the failure sequence is neither checked nor
allowed for; and for this reason it is sufficient to limit the member strength
and stress-strain calculations to the prescribed elements only. A failure
sequence analyzed in this way is only defined conditional on the survival of
6-37

all the other elements of the structure and does not have the same physical
interpretation as the failure sequence produced by a deterministic collapse
analysis. However, since all possible failure sequences are, in principle,
considered, the union over these conditional failure sequence events gives the
system failure probability which is the final objective of the analysis.

(c) A single structural member may have more than one failure mode and the
member capacity is determined by the most critical of these modes. For a
given set of resistance and load variables a deterministic analysis will identify
the critical failure mode and determine the member capacity accordingly.
However each failure mode is influenced by different variables which will, in
general, have different probabilistic descriptions. In a reliability analysis it is
necessary to evaluate the probability of member failure in each mode
separately. Although the probability of an initial failure of a member in a
particular failure mode may be marginally low, for example brittle fracture, the
probability of occurrence of a failure sequence initiated by such a mode could
be higher than that involving any other set of member failure modes.

In a reliability analysis, for reasons similar to those mentioned in (b) the VDM
algorithm is modified to introduce state-changes for each element on the
failure sequence in the precise mode prescribed while the limit exceedance of
other modes for the same member is not checked. Thus it is more appropriate
to refer to each component of the prescribed failure sequence as a 'failure
element' and not as a structural element.

The modified VDM algorithm to be used within an iterative search for the joint
beta-point for the reliability analysis of a prescribed failure sequence can be
summarized as below.

(i) Choose a starting solution of variables u=u0 from previous analysis or set
u={O}. Set a flag for identification of convergence, IBET=O. Compute the
distance from the origin, 1= u
6-38

(ii) Transform the current iterate u from standard normal space to x in the space
of basic variables.
(iii) Initialize the set A of failed elements A=(0). Initialize total stresses , total
strains , virtual distortions and incremental virtual distortions M. Initialize
logical flag for collapse condition, COLL=.FALSE.
(iv) Compute random linear response to environmental loading using the response
surface approach, Eq.4.16 and Eq.4.17.
(v) Determine capacity of the (next) element on the prescribed failure sequence for
the prescribed failure mode.
(vi) Calculate safety margin (constraint function value) for the prescribed
limit-state, Eq.6.26. If M^O go to step (xii). If j=q (i.e. last component on
the failure sequence) and IBET^1 (i.e. no convergence yet) go to (xiii).
(vii) Calculate overstress coefficient ' (see Gierlinski and Lau, 1991) and include
elementj in the set A=Au(j).
(viii) Adjust internal forces of elements in set A to follow their post-limit behaviour
(i.e. hardening/softening corrections, flow rule etc.)
(ix) Accumulate current distortions , calculate surplus stresses Ic and increment
of virtual distortions M (Eq.6.8) from all failed elements.
(x) If the matrix B is singular, structural collapse occurs. Then set flag
COLL=.TRUE. and go to step (xiii).
(xi) Determine current total strains and stresses from Eq.6.7.
(xii) If j=q (i.e. last component on the sequence) go to step (xiii) else return to
step (v) to examine the next element on the failure sequence.
(xiii) If IBET^1, repeat steps (ii) to (xii) n times to compute gradients of constraint
functions. If IBET=1 (i.e. solution has converged) go to step (xv).
(xiv) Determine d= UI. If (d-d0 )^tolerance, then joint beta-point found!. Set u=ü,
IBET=l and go to step (ii). Else, set d0 =d, generate new iterate and go to (ii).
(xv) If IBET=1 then calculate failure probability p using Eq.6.25, write results to
data-base. If (COLL) then the sequence is termed complete failure
sequence' else an 'incomplete failure sequence'. Take up a new failure
sequence and go to step (i). If no more sequences, STOP.
6-39

The above algorithm has been implemented by the author in the RASOS software
which enables the treatment of more rigorous softening/hardening behaviour for
primary members and tubular joint failures by plastic collapse or fatigue-fracture, see
Shetty and Baker (1991d). With this approach it is possible to analyze failure
sequences which involve different failure modes for the same structural member.
Thus, for example, a tubular joint failure at a member end followed by a buckling
failure of the member can be studied. Each component in a failure sequence thus
refers to a 'failure element' and not necessarily to a structural member.

The RASOS software enables the user to choose from a total of 9 limit-state criteria
for a member/joint, as listed in Appendix-B, and more than one limit-state type could
be prescribed at a member end, even within the same failure sequence, see Shetty and
Baker (1991d). A total of 60 basic variable types can be used for uncertainty
modelling, as listed in Appendix-C. Thus the RASOS software enables to perform a
more rigorous and yet practicable reliability assessment of jacket structures.

6.6 SYSTEM RELIABILITY ANALYSIS

In the previous section the reliability evaluation for a single (prescribed) failure
sequence was considered. However, for a large redundant structure the number of
possible failure mechanisms/paths will be quite large and structural failure may result
from any one of these modes. The system failure event is then defined as the union
of all modal failure events and the system failure probability is evaluated as

(6.27)
pf,(YS - P[UF]

where, Fk denotes the event "failure in the kth sequence/ mechanism". It would be
practically impossible to identify all the N failure paths of a large structure and
usually only a small number of failure paths contribute significantly to the system
failure probability. One of the major tasks in the evaluation of structural system
reliability is the identification of these, so called, "dominant failure paths". A number
6-40
of methods have been proposed for this purpose in the literature which can be grouped
as collapse mechanism based methods and failure-tree enumeration methods. These
methods are briefly reviewed in sections 6.6.1 and 6.6.2 respectively. A new method
of failure-tree enumeration which is suitable for analysis of large structures is
developed in section 6.6.3.

6.6.1 Identification of Dominant Collapse Mechanisms:

One of the approaches to identify a set of dominant failure mechanisms is to use the
fundamental mechanisms for the structure, Thoft-Christensen and Murotsu (1986).
This is in line with the plastic analysis method discussed in section 6.3.1 and the
reliability analysis method for a single failure mechanism given in section 6.5.1
before. In this method a set of fundamental mechanisms are first identified for the
structure by the method of Watwood (1979). Then reliability analysis is carried out
for each of these mechanisms using FORM based on a safety margin expression of
the type given by Eq.6.20. The next step is to select a number of fundamental
mechanisms as starting points for forming new mechanisms. For this purpose the real
fundamental mechanism with the lowest reliability index is identified and all real
fundamental mechanisms in the interval where t is a pre-specified
constant, are selected. Let (3l^I32^...^3f be an ordered set of f fundamental
mechanisms selected as starting points. Next the fundamental mechanism 1 is
combined with the all the remaining (rn-i) fundamental mechanisms 2,3,...,m and
reliability indices f312'13'•'f31,m are determined for the combined mechanisms. The
mechanism with the smallest reliability index identified and the new mechanisms with
reliability indices within a range of A 2 from the smallest are selected for further
investigation. The same procedure is then used for the fundamental mechanisms
2,3,... ,f to identify sets of combined mechanisms with relatively lower reliability
indices. The procedure is continued,-each -time adding or subtracting a fundamental
mechanism to the previously selected combined mechanism and discarding combined
mechanisms with very high reliability indices, until a reasonable number of combined
mechanisms with relatively high probabilities of failure are identified. Finally all the
6-41

significant mechanisms are used as elements of a series system to calculate the system
probability of failure. Since only a sub-set of all possible mechanisms are used the
method results in a lower bound to the correct probability of failure (or upper bound
on system reliability).

A method which does not require the generation of fundamental mechanisms has been
given by Ditlevsen and Bjerager (1984), see also Hoim et al (1988) and Baker (1985).
In this method a lower-bound on system reliability is computed using the static
theorem (or stable or lower-bound) of plasticity. This states that the structure will be
able to carry the external load if and only if there exists a statically admissible set of
internal forces such that these nowhere violate the yield condition, Heyman (1971).
Consider an indeterminate structure, with m possible hinge locations in which bending
moments are dominant. Let the structure be acted upon by a set Q of external loads
of random magnitude. Then, by choosing a statically determinate primary system such
that the bending moments Sk at n of the possible m hinge locations are assigned
values z=(z,z2,...,zj, then 'lower-bound safety margins' for the m locations can be
defined as

Mk - Rk - k-1,2,...,n
(3.28)
1

Mk - Rk-. EakJQJ + k-n+1,n^2,...m


j-1 i-i

where Rk is the plastic moment capacity of section k. The system reliability is then
bounded by

P[(M1>O)fl(M2>O)fl. •fl(Mm>O)1 (6.29a)

or identically


> P[min(Mk>O)] (6.29b)
6-42

The best (highest) lower-bound on system reliability can now be found by selecting
the arbitrary values of the n moments z so as to maximize the right-hand side of
Eq.6.29. For normally distributed random variables and deterministic dimensions this
can be carried out using linear programming methods while for more general cases
non-linear programming is required, Baker (1985). Often a near optimal solution can
be determined by maximizing the reliability for the component with lowest .

However, the lower-bound on determined in this way turns out to be much


smaller than the true value. Moreover, the results depend on the choice of the primary
redundant system. Ditlevsen and Bjerager (1984) also give a method for obtaining an
upper-bound on system reliability (lower-bound on A linear combination of
lower-bound safety margins gives

(6.30)
M - EYkMk

If M is independent of z and all 'Yk are non-negative it can be shown that M is an


'upper-bound safety margin' corresponding to a plastic mechanism. By combining
a dominant lower-bound safety margins a number of dominant upper-bound safety
margins can be formed. The final system reliability is evaluated as a series system
of upper-bound safety margins.

In both the methods described above it is not possible to guarantee that none of the
significant mechanisms have been missed. The methods identify dominant
mechanisms in arbitrary order and sometimes the same mechanism may be identified
more than once.

Perhaps the most rigorous of the proposed mechanism based methods is the 'polytope
extreme point' (PEP) method given by Nafday et al (1987). The method is based on
a kinematic-static pair of linear programming models. The optimal solution of the
kinematic (primal) LP gives the critical collapse mechanism (deterministically) as a
linear combination of member resistance or as a linear combination of fundamental
6-43
mechanisms. The static (dual) LP model, on the other hand, gives the collapse load
and the bending moment distribution for the critical collapse mode.

The feasible region defmed by the constraints of the primal LP (see Nalday for
detailed equations) represents a convex polytope (polyhedra) while the objective
function represents a hyperplane. A feasible solution of the primal LP is seen to
correspond to one of the extreme points of the polytope. Thus the problem of finding
all the failure mechanisms turns out to be a problem of finding all extreme points of
a polytope for which a number of algorithms are available in the relevant
mathematical literature. In the proposed method by Nafday et al (1987) the failure
mechanisms are identified in the ascending order of their collapse loads. The method
automatically excludes all kinematically inadmissible failure modes and can generate
a certain prespecified number of failure modes or all mechanisms up to a prespecified
level of significance. The method has been demonstrated on a simple example.

6.6.2 Failure-Tree Enumeration Methods:

Most of the reported studies on system reliability of structures have used a failure-tree
enumeration approach. In this approach a sequence of component failure events
leading to the collapse of the structure is traced. Since after each elemental failure
(called a 'node') all the surviving elements can be chosen to form the next
component to fail, the failure events can be represented in the form of a failure-tree
with each branch of the tree representing a failure sequence. Of the methods proposed
to develop a failure-tree with stochastically dominant failure sequences, the
'incremental load method' of Moses (1982), the 'branch and bound method' of
Murotsu et al (1984), the 'truncated enumeration method' of Melchers and
Tang (1984) and the method of Guenard (1984) have found wide acceptance.
Although the methods have been proposed in relation to either simple truss or frame
formulations, the failure-tree enumeration approach used in these studies is fairly
independent of material/member behaviour and the method used to calculate failure
probabilities of individual failure sequences. The incremental load method of
6-44
Moses (1982) does not have a systematic enumeration strategy and is not discussed
further. Fundamental to all the remaining methods is: (a) a branching criterion, (b)
a truncation criterion and (c) a criterion to determine structural failure. The above
methods differ only slightly in each of these aspects. The basic features of only the
branch and bound method given by Murotsu et a! (1984) is reviewed below, see also
Thoft-Christensen and Murotsu (1986), as this was the earliest method to be proposed
and has been widely used in reported studies on system reliability.

(a) Branching Criterion:


The starting point in a failure-tree enumeration method is the intact structure in which
none of the elements have failed. This is denoted state '0' or 'initial' state.
Reliability analysis is carried out for all the elements (all failure modes of all
structural elements) of the structure assuming no other elements fail. This is termed
a 'Level-O' analysis and the reliability indices as 'Level-O reliability indices'.

Next, one of the elements has to be selected as a starting node for further branching.
It is logical to select the element with the largest probability of failure or the lowest
reliability index as this will often, though not always, lead to the most likely failure
sequence. Starting with this node, failure sequences with two elemental failures are
formed by combining in turn all the remaining elements of the structure. The
probability of failure of these pairs of elements is determined and the sequences are
arranged in a descending order of their probability of failure and the branch with the
highest probability is chosen for further branching at Level 3.

Thus at any level j, the branch with the highest path probability is chosen. The path
is first examined to see if it results in structural failure. If a path results in system
failure it is called a 'complete failure path' or else a 'partial' or 'incomplete failure
path'. A selected partial failure path is branched one level higher (j+1) by combining
one element in turn out of all the surviving elements (i.e. the elements which are not
already present on the chosen path). A partial failure path, once chosen at Level j is
pursued to higher levels tracing all its branches before returning to examine the next
6-45

partial failure path at the same Level j . Thus none of the branches are truncated until
the first complete failure path is found.

(b) Truncation Criterion:


As mentioned earlier, not all branches of the failure-tree will be of same importance
in computing the system failure probability p 15,. A truncation criterion by which a
kth complete failure sequence with probability p could be ignored is given as, see
Meichers and Tang (1984)

Pf,k
&:. ^5 (6.31)
p!,:,3 ;

where S is a small constant expressing the degree of accuracy desired in computing


pf, and thus p S gives the threshold failure probability. Since the system
probability of failure Pf,s is not known at the outset it is approximated by the modal
failure probability Pf, where

k-I
- max(p1, ,) ^ p1,,,3
(6.32)

for the k-i previously identified complete failure paths. Use of p1 (^ f ) in Eq.6.31
is obviously conservative, and may allow a failure mode k to be accepted rather than
rejected if p f had been available, Meichers and Tang (1984). Similarly a truncation
criterion for an incomplete failure path at Level j can be derived by using the path
probability for p in Eq.6.31. Since the failure path probability monotonically
decreases as the path length increases, further branches from a partial failure path for
which the probability is already below the threshold value will only result in
insignificant failure sequences.

Once the first complete failure sequence is found at Level q, all complete and
incomplete failure paths at this level with failure probabilities below the threshold
level are truncated. If no partial failure paths are left at this level for further
branching, the attention is switched to the next lower level q-i. At this level again
all branches with probabilities below the threshold level are discarded and the path
6-46

with the highest probability is chosen for further branching. This procedure is
continued, working successively backwards and forwards, until all partial failure paths
are either discarded by the truncation criterion or traced up to system failure.

(c) S stem Failure Criterion:


y

In all the above mentioned methods the system failure is defined as the collapse of the
structure by formation of a mechanism. The structural failure is identified by
monitoring the determinant of the structural stiffness matrix, with K I^O denoting
collapse. A failure path which results in structural collapse is called a complete
failure path and further branching from this path is terminated while a partial failure
path, which is not discarded by the truncation criterion, is taken up for further
branching.

6.6.3 A Practical Approach to Failure-Tree Enumeration:

The branch and bound method presented above is theoretically rigorous and can be
used to identify all the dominant failure paths within a chosen tolerance. The
truncation criterion in Eq.6.31 combined with exhaustive enumeration at each level
ensures that none of the dominant failure modes will be missed. However, the method
can be computationally very expensive for large structures. A more practical oriented
failure-tree enumeration approach is developed in the following which is feasible for
analysis of large structures. The approach is in line with the reliability analysis
method based on joint beta-point concept proposed in sections 6.5.3 and 6.5.4. The
overall approach as presented below is that of the author but it uses some of the ideas
from previous works, namely Guenard (1984) and Gollwitzer (1991), and these are
explicitly acknowledged.

Selection of Failure Elements:


In a large structure, such as an offshore jacket, there will be a large number of
members each with multiple failure modes. However, a number of members are
designed for conditions of fabrication, load out, transportation, lift or launch and these
6-47

elements may not be critical for the in-place condition for which a reliability analysis
is normally carried out. Thus initial selection of failure elements for use in a
failure-tree analysis may be based on the critical members and dominant failure modes
for each member obtained from a linear-elastic deterministic analysis. However, a
number of elements which may be lightly stressed in the intact state may become
overstressed in a damaged state during load-effect redistribution. Thus additional
elements based on contiguity with critical members, availability of alternate load paths
etc. have to be included. Deterministic progressive collapse analysis for various wave
directions should be carried out to obtain detailed understanding of the load-shedding
behaviour of the structure.

System Failure Criterion:


In the branch and bound method structural collapse, identified by 1K I ^0, is used to
defme system failure. From the analysis of realistic full-scale jacket structures, see
Example 1 in this chapter and Sears and Shetty (1992), it has been observed that a
failure path up to collapse may involve a large number, typically 15-20, of elements
(plastic hinges). However, no appreciable change in load carrying capacity is
observed after the failure of typically 8-12 elements. Thus for practical purposes a
failure criterion less than complete collapse could be used. With the virtual distortion
method of non-linear analysis, in which the stiffness matrix is not modified, the
system failure criterion may be stated in terms of the active influence matrix as
JB I^. Unlike the stiffness matrix, the determinant of B matrix has the property of
being lB =1 for the intact structure and lB =0 corresponding to collapse and thus
provides a relative measure of residual strength of the structure. The constant may
be selected based on deterministic collapse analysis of the structure for various wave
directions. Alternatively, system failure based on a limiting value of deflection of the
top deck from operational considerations could be used.

Initial Estimate of Modal Probability of Failure p:


The truncation criterion given by Eq.6.31 will obviously be more efficient if the value
of p corresponds to the probabilistically 'most dominant complete failure path'
since in this case more non-dominant modes will be eliminated. In the branch and
6-48

bound method, initially the branching is continued, choosing at each level the branch
with the highest probability, until the first complete failure path is found. The Pii
value for this path is then used as the initial estimate of p for truncation. This
branching procedure, although identifies one of the dominant modes, does not
necessarily lead to the most dominant mode, Meichers and Tang (1984). Moreover,
none of the paths could be truncated by this procedure until the first complete failure
path is found.

To avoid this situation, it is proposed that a reliability analysis be first carried out
directly for the most critical failure sequence obtained from a deterministic progressive
collapse analysis of the structure for various wave directions. When the uncertainty
in member capacities is small (COV<O.30) compared with that of the common loading
and the resistance variables are highly correlated all over the structure, experience
shows that the dominant failure sequence obtained from a deterministic collapse
analysis will be identical to the probabilistically most dominant failure path. These
conditions are often met in practice for jacket structures and the failure path
probability so obtained should be a good choice for p. This can, however, be revised
if a higher value is found during failure-tree analysis.

Enumeration and Truncation:


In the branch and bound method the enumeration is exhaustive i.e. a selected branch
at Level (j-1) is branched to the next Level j by combining the surviving elements,
one at a time, and calculating failure path probabilities for all the newly formed
branches at Level j before any branches are truncated. Note that the order in which
the new elements are combined is completely arbitrary. In a real structure with a
large number of failure elements the calculation of these partial failure path
probabilities at each level can be laborious. Moreover, many of these branches are
eventually discarded by the truncation criterion thus wasting expensive calculations.

In the proposed approach, at a Level (j-l), a deterministic analysis is carried out for
all the surviving elements to calculate their safety margin values and the elements are
arranged in the ascending order of their safety margins. These calculations are carried
6-49

out at the joint beta-point of the selected partial failure path at Level (j-l). Next,
starting with the top-most element which has the lowest safety margin, branches axe
formed by combining each element and calculating the path probability at Level j.
Each branch so formed is checked for truncation using Eq.6.31 (using the p value
which is already available) and further branchings are stopped once any branch meets
the truncation criterion. Thus at Level 0, for example, safety margins are calculated
for all failure elements at the mean values of basic variables and the elements are
arranged in decreasing order of their criticality. Starting with the most critical element
failure probabilities are calculated for each element until the probability for an element
falls below the threshold value.

As discussed earlier, when the loading uncertainty dominates, the ordering of failure
elements based on safety margins (a deterministic measure) will be identical to the
ordering based on reliability indices. Thus once the path probability during branching
from a chosen partial failure path falls below the threshold level further branchings
could be stopped. In the proposed approach, therefore, enumeration is selective and
truncation is applied at each level and after branching from each chosen failure path.

Although the idea of using safety margins for sorting the components was first
proposed by Gollwitzer (1991), it was not used as a basis for truncation in the manner
proposed as above.

Branching Criterion:
Once branchings from a failure path are stopped, a next partial failure path is to be
chosen for further branching. For this purpose, unlike the branch and bound method,
partial failure path which has the highest path probability at that stage is chosen
regardless of its level as proposed by Guenard (1984). This criterion for selection of
a branch k can be stated as

Pf,k - max(P,[E1 flE2 fl...flE]) for O^j^q (6.33)


6-50

where q1 is the length of the kth complete failure path. Thus, with the provision to
jump between levels and always branching from the path with the current highest path
probability, the first complete failure path found will be the probabilistically most
dominant failure path. Moreover, additional complete failure paths will be identified
in decreasing order of their failure probabilities. This feature has a further useful
outcome as discussed next.

Criterion for Terminating the Analysis:


In the branch and bound method the failure-tree enumeration is continued until all
failure paths are closed either due to truncation or they result in system failure. For
a large structure, even with the efficient enumeration and truncation strategy proposed
above, it will be too difficult to reach this stage and an alternative criterion to stop the
analysis at an earlier stage is desirable. For this purpose a criterion based on
bounding the system failure probability is proposed below.

Although it is difficult to obtain the exact system probability of failure as in Eq.6.27


bounds on Pf,sys of varying closeness can be derived easily. Within a specified level
of significance (ö in Eq.6.31) a set of crude and best bounds on Pf,s can be given as:

Crude upper bound:


(6.34)
P1, UB - [
E]

Crude lower bound:



Pfs,i. - PDC(ElflE2fl...flEqDC) (6.35)
6-51

Best upper bound:

(mm
(6.36)
Pf,uB - P1 U [ElflE2fl...flE(q.k)l]]
k-1

Best lower bound:

(Rn
(6.37)
- Pl U [ElflE2fl...flEq.k]]
(k-1

where E1 are individual elemental failure events and m is the maximum number of
failure elements included in the analysis. The crude lower-bound corresponds to the
probability of the complete failure path which is identified by a deterministic collapse
analysis. In evaluating the best upper-bound the union is over all mm dominant
partial failure paths, each traced up to one level below system failure while in the case
of the best lower bound the union is over all nn dominant complete failure paths.

The branch and bound method seeks to determine the best lower-bound but it does not
give any indication of how close it is to the exact Pf,3y5. Bounds of intermediate
closeness can be obtained if m'^ mm failure paths are traced up to Tk ^ (q-l) levels
for use in Eq.6.36 and n^ nn dominant complete failure paths are identified for use
in Eq.6.37. Since in the proposed method the failure paths are enumerated in the
order of their dominance the bounds become successively closer as the analysis
progresses. The bounds can be evaluated at certain intervals, after a reasonable
number of further branchings, and the analysis can be stopped once the bounds
become acceptably close.

Because of the various approximations used it can be argued that the proposed method
does not guarantee that none of the dominant failure modes will be missed. However,
the final objective is to estimate the system failure probability and the efficacy of the
method should be judged on how efficiently bounds can be derived for which are
sufficiently narrow. The feasibility of the proposed method for analysis of large
6-52

structures should be evaluated by carrying out a number of realistic case studies and
comparing with the more rigorous but expensive branch and bound method.

While the full algorithm is yet to be implemented into a computer code and validated,,
a number of ideas presented above have already been incorporated in the computer
program RASOS_R, see Shetty et al (1991) and Gollwitzer (1991). The program
offers both interactive and automatic failure-tree analysis with a facility to STOP and
RESTART the analysis at any stage.

Finally a note on the calculation of system failure probabilities typically expressed in


the form of upper or lower bounds as in Eqs.6.36 or 6.37. This is called a cut-set
representation in which parallel sub-systems (cut-sets) are arranged in series. In
principle two methods are available for the calculation of failure probabilities of
cut-set systems as discussed in detail in chapter 2 (section 2.6.3). In summary, the
first option is to use Ditlevsen bounds as given by Eq.2.51. The required intersections
(FrF) between pairs of components, each of which is now a parallel system with m1
and mj number of elements, is evaluated using a (m 1+m)-dimensional standard normal
integral. In the second method, each parallel system, the probability of which is
calculated using multi-normal integral as in Eq.6.25, is replaced by an equivalent
hyperplane of the same reliability. The final union made up of equivalent components
is evaluated as for pure series systems, again using the multi-normal integral as in
Eq.2.48. The second method is used in RASOS_R program as it is more robust even
for ill-conditioned systems and directly computes a first-order estimate of the system
probability of failure instead of bounds as in the first method.

6.7 RELIABILITY ANALYSIS UNDER FATIGUE DETERIORATION

In the discussion in previous sections no consideration was given to the effect of


fatigue damage on the reliability of a structure. The proposed reliability analysis
method is essentially a time-independent formulation based on random variable models
for description of uncertainties in basic variables. The methodology developed so far
6-53
can thus be used for reliability analysis of a jacket structure under a single extreme
wave event in which the structure fails essentially by failure of tubular members.
Ignoring any form of strength deterioration in time and assuming that wave loading
is dominant, the life-time reliability of a jacket structure can be approximated by the
system reliability under the life-time extreme wave condition. In this type of analysis
the wave height is described in terms of its life-time extreme distribution often using
a Gumbel model. The other environmental parameters such as wave period, current
and wind are described in terms of conditional distributions for a given height or more
simply in terms of marginal distributions and some form of correlation structure
between the variables. Most reported studies on system reliability analysis of jacket
structures have adopted this approach, see Baker (1985) for a detailed review.

The above approach, however, will not be valid when the effect of fatigue is to be
accounted for. As mentioned earlier, most reported studies are based on a member
replacement method of reliability analysis. In this approach, more than one failure
mode for a single structural member cannot be easily considered. For this reason
reliability analyses were carried out separately for the extreme wave condition,
considering ih tubular member failures, and for the fatigue condition, considering
jj fatigue failure of tubular joints. However, recent experiments on large tubular
frames indicate that joint flexibility and strength can have a significant influence on
the ultimate behaviour of frames, Lalani and Shuttleworth (1990). Therefore, even
when the two conditions are treated separately it is important to include tubular joint
failures due to plastic collapse or fracture failure from significant weld defects in an
extreme wave reliability analysis. At the present time no studies have been reported
which account for both tubular member and tubular joint failures in a system
reliability analysis. Extension of the methodology developed in the previous sections
to include tubular joint failures and fatigue deterioration is considered in the following.
6-54

6.7.1 Incorporation of Plastic Collapse and Fracture of Tubular Joints:

The proposed reliability analysis methodology uses the modified virtual distortion
method of non-linear analysis as given in section 6.5.4. With this method more than
one limit-state for each structural member can be explicitly treated. Failure sequences
can now be prescribed with each failure element representing a member failure
(yielding/buckling) or a joint failure (plastic collapse/fracture). It is also possible to
have a joint failure by plastic collapse at a member end followed by buckling of the
same member, however, a fracture failure of a joint cannot be followed by any other
failure elements for the same member for obvious physical reasons. The use of the
joint beta-point concept and multi-constrained optimization further helps to ensure that
the non-linear analysis does not stray into physically inadmissible regions of the
probability space.

A safety margin equation (constraint function) for plastic collapse of a joint,


following (j- 1) member or joint failures, can be written as in Eq.5.9 of chapter 5. The
brace internal forces pW, M1 and M0 for the Jth failure element should be calculated
after (j-1) successive load-effect re-distribution following each elemental failures. The
elemental state-changes are simulated as virtual distortions and modified stress/strain
state of the structure is calculated using the virtual distortion method. The joint forces
will now be functions of random environmental parameters, which influence linear
response, and the resistances of (j-l) failed elements, which influence non-linear
response of the structure. Following plastic collapse failure of a joint its residual
capacities are set at their limit values P, M1 and Mo representing a plastic post-limit
behaviour as discussed in section 6.2.2 and the surplus forces are re-distributed to the
remainder of the structure.

Similarly, a safety margin equation for fracture failure of a joint can be written as in
Eq.5.l 1 of chapter 5, using the CEGB R6 approach (or as in Eq.5.16 using the
BS PD 6493 approach). In this case the brace internal forces determined as above are
used to calculate the plastic collapse parameter L 1 (or ST) and hence (or 5). The
6-55

hot-spot stress S determined from these forces and the residual stress S are used
to calculate the fracture parameter K (or ö) corresponding to a weld defect of random
initial size (a0,c0). Following fracture failure of a joint its residual capacity is
conservatively set to zero representing a brittle post-limit behaviour as discussed in
section 6.2.2 and the surplus forces are re-distributed to the remainder of the structure.

6.7.2 Incorporation of Fatigue Deterioration:

If the effect of fatigue is not considered, the two joint limit-states, treated as above,
combined with other member limit-states can still be used in conjunction with a
life-time extreme wave height distribution to estimate the life-time reliability of the
structure. However, when fatigue effects are present the problem becomes more
complex and the approach presented above has to be modified as discussed next.

The effect of fatigue can be treated in a system reliability analysis with varying
degrees of complexity. The system reliability method developed in the previous
sections is essentially a time-independent formulation in which fl elemental failures
are considered to occur under a single (monotonically increasing) load event. The
fatigue failure (through-thickness crack formation) of tubular joints, on the other hand,
is essentially a time-dependent phenomenon, in which successive joints can fail at
different (random) points in time. Such a scenario is more difficult to accommodate
in a time-independent formulation. In this section, therefore, a simplified approach
is used in which the effect of fatigue is included mainly as a deterioration in strength
of tubular joints under plastic collapse and fracture failure modes. A more rigorous
treatment of fatigue effects, which also explicitly considers the fatigue limit-state is
given in chapter 7 where a time-dependent system reliability formulation is developed.

The capacity of a tubular joint under plastic collapse mode decreases if a crack of
significant size is present at the joint. The strength of cracked joints can be calculated
as presented in section 4.2.6. Similarly the propensity to fracture failure increases as
the crack size increases due to fatigue. If the service exposure period is specified as
6-56
t years, the random crack size a(t) at the end of time t can be calculated as a function
of random initial defect size and other random variables which affect fatigue crack
growth as discussed in chapter 5 (section 5.8.1). The safety margins for plastic
collapse and joint fracture in the previous section can now be evaluated as for a
cracked joint with a fatigue crack of size a(t). In this case all fatigue random
variables for all joints should be included in the total list of basic random variables
for the structure. The rest of the reliability analysis procedure remains unchanged.

In the above approach for including fatigue effects, it is to be noted that the extreme
wave event and structural failure is considered to occur at any time within the t" year
while the fatigue deterioration is conservatively calculated to the end of the tth year.
Therefore the wave height should be described in terms of its annual extreme
distribution and the other environmental parameters with their conditional/marginal
distributions associated with the annual extreme wave height. The computed system
probability of failure will therefore correspond to an annual failure rate. Since the
fatigue deterioration depends on time t, system reliability analysis has to be performed
for various durations of specified exposure to study the variation of system reliability
through the service life of the structure. In this approach, therefore, life-time
reliability of the structure cannot be computed when fatigue effects are included.
However, the annual probability of failure can still be used as a rational measure in
the design of tubular joints, in planning in-service inspection intervals or for decisions
on repair of detected fatigue cracks.

6.7.3 Choice of Failure Elements Included in Failure-Tree Enumeration:

As can be expected the search for dominant failure paths constitutes the most
expensive part of any system reliability analysis. Every effort should be made to keep
the -computational effort to a minimum for the -analysis to be practicable for large—
structures. A judicious selection of member and joint limit-states and choice of basic
variables is important in this respect. The behaviour of tubular members can be
modelled using simple code formulations or using more sophisticated non-linear
6-57
beam-column elements. Similarly, the fatigue-fracture behaviour of tubular joints can
be analyzed to varying degrees of complexity. The RASOS_R program, for example,
models each member end as a basic structural unit and for each member end more
than one limit-states can be assigned from out of a total of 4 member limit-states and
5 joint limit-states as given in Appendix-B. A total of 60 variable types are available,
as given in Appendix-C, for modelling uncertainties in loading and resistance
properties depending on the limit-state of interest.

During failure-tree enumeration, it may be sufficient to use simple and inexpensive


code formulations to analyze yielding and buckling failure of tubular members. Strain
fracture of tubulars is unlikely for the type of steel used in offshore construction and
may occur only after large plastic deformations. Moreover, yield strength of steel in
tension and compression may be assumed to be fully correlated and consequently only
one variable per member end can be used to model yield stress variability. However,
full correlation may not be appropriate for yield strength values at the ends of a
member. Once all dominant failure sequences are identified failure probabilities for
important failure sequences can be recomputed using more sophisticated member
formulations and uncertainty models.

The problem is more complex in the case of tubular joints, especially when expensive
fracture mechanics fatigue crack growth calculations are included within iterative
reliability analyses. Experience from example analyses shows that for member
limit-states and for the limit-state of joint collapse the reliability index is more
sensitive to variability in extreme environmental loading parameters while the tubular
joint fracture limit-state is dominated by uncertainties in resistance variables. This
causes convergence problems during the search for joint beta-point when relatively
unlikely failure sequences involving combination of member and joint limit-states are
being analyzed. The problem will be compounded when fatigue effects are included.

To minimize such problems it has been found helpful to perform the analysis in two
stages. In the first stage, a detailed study on reliability of isolated tubular joints is
carried out for all tubular joints in the structure (or selected critical joints from
6-58

deterministic analyses) to establish relative dominance of the limit-states of fatigue,


plastic collapse and fracture for each joint. Experience with real structures, see
Example 1 in the following, suggests that for most joints only one of these limit-states
dominates. Only for few joints interaction of fatigue and fracture or fracture and
plastic collapse is seen to be important. From the results of example analyses in
chapter 5, it can be seen that in most cases the plastic collapse limit-state is not
influenced by fatigue deterioration. For the high toughness steels used in offshore
construction it is seen that the fracture behaviour is dominant for as-welded joints with
high tensile residual stresses while the plastic collapse limit-state dominates in the case
of stress-relieved joints. Based on this study all tubular joints can be categorised into
"fatigue-critical", "collapse-critical", "fracture-critical" or "mixed behaviour" groups.
In the second stage of system reliability analysis, each joint group is assigned its
corresponding dominant limit-state and only those basic variables which are relevant
for a particular limit-state are used for modelling uncertainty in resistance of each
joint. A set of important basic variables have been identified for each limit-state and
reported in section 5.9 of chapter 5. Again simple and inexpensive formulations can
be used during failure-tree enumeration and probabilities of important sequences
recomputed using more sophisticated models. If necessary, additional failure
sequences around the important sequences can be formed and analyzed using
alternative limit-states for the same joints.

Example 1: Reliability Analysis of a Jacket Platform Under Extreme Wave

Reliability analysis of a full-scale jacket platform using the proposed methodology and
the RASOS software is presented in the following The structure considered is a
6-legged jacket platform, called "BRITE", standing in a water depth of 60 m. The
structural geometry and loading conditions are described in Appendix-D. The analysis
is carried Out in three main stages as detailed below.
6-59

PRELIMINARY LINEAR ELASTIC ANALYSIS

The jacket structure is idealized as a space frame with the tubular members modelled
as beam-column elements. The computer model of the structure is generated using
the program module RASOS_M, see Appendix-D for detailed plots.

The environmental loading on the structure is generated using the module RASOS_E.
The loading can be expressed as distributed member loads or equivalent nodal loads.
For a deterministic design assessment of the structure the environmental parameters
as given in Appendix-D are used while for reliability analysis the loading is
determined based on mean-value environmental parameters. In addition to the total
load, the program also generates 8 deterministic components of environmental loading
for later use in response surface load modelling during reliability analysis. Only one
wave approach direction (along +ve Y-axis) is considered.

A linear-elastic static analysis of the structure is carried out using the module
RASOS_L. Gravity loading due to self-weight of the jacket, deck loads applied as
concentrated loads at the top of the jacket and environmental loading are analysed as
separate load cases. Structural response in the form of nodal displacements and
member internal forces are determined. Response is also determined separately for
the 8 deterministic components of environmental loading.

The results of the linear analysis under gravity and design environmental loading
(wave height = 22m) show that most of the members are highly under utilized. A
summary of highest utilization ratios (load/resistance) for different groups of members
at different elevations in the structure are given in Table 6.1. From this table the
following important points can be observed:
6-60

Table 6.1: Highest utilization ratios for different groups of members.

Y-Frame Y-Frame X-Frame X-Frame


Level Legs
VDG HZB VDG HZB
1 0.35 0.02 0.11 0.01 0.32
2 0.37 0.07 0.06 0.04 0.25
3 0.33 0.10 0.07 0.06 0.17
4 0.28 0.08 0.06 0.04 0.10

VDG: Vertical DiaGonals


HZB: HoriZontal Braces
X-Frame: A frame parallel to X-axis
Y-Frame: A frame parallel to Y-axis

* Only the jacket legs and vertical diagonals are highly stressed compared to
other members. The horizontal braces and the horizontal diagonals are the
least stressed members in the structure.

* The vertical diagonals in the two frames parallel to the Y-axis resist most of
the wave loading while the X-direction frames carry relatively little load.

* A further examination of results shows that corresponding members in the two


Y-frames carry equal loads due to the symmetry of the structure and loading.

* Although the compressive / tensile pair of diagonals at a level carry the same
stresses, the utilization ratios for compression members is relatively high
because of the lower buckling capacities.

* The two compressive diagonals in Level 2 (Elem. Nos. 2134 and 2336) are the
most highly stressed members and may be expected to fail first if the
horizontal loading is increased.
6-61

DETERMINISTIC PROGRESSIVE COLLAPSE ANALYSIS

The structural influence matrix (fl-matrix) required for non-linear analysis is first
generated using the program module RASOS_D. This program uses the stiffness
matrix generated by the RASOS_L program to develop the D-matrix using the virtual
distortion method. A progressive collapse analysis of the structure is then performed
using the RASOS_C program. The required member limit-forces under axial tension
and bending moments are determined based on nominal cross-section area and plastic
section modulus and mean value of yield stress in tension (300 MPa). The limit
forces under axial compression and post-limit slope (softening) are determined using
a 3-D elasto-plastic buckling analysis for each member. A nominal post-limit
hardening slope of 0.02 is assumed under axial tension. A limit-surface is developed
to account for the interaction of axial load and bending moments.

The progressive collapse analysis is performed by increasing the loading on the


structure in several increments. In the first increment, all the gravity load, deck load
and design environmental loading is applied so that the results correspond to a linear
elastic analysis. Next only the horizontal environmental loading is increased by
scaling the design environmental loading by a load factor. During the loading process,
when the limit-surface at a member end is exceeded (formation of a plastic hinge), the
surplus force from that hinge is re-distributed to the remaining members using the
virtual distortion technique. The forces in the failed element are adjusted to follow
the post-limit behaviour of the member. The process is repeated for every new plastic
hinge formed during loading and the load incrementation is continued until the
structure collapses by the formation of a mechanism.

The sequence of plastic hinges and the corresponding load levels are shown in
Table 6.2. The determinant of the active influence matrix based on failed elements
(B-matrix) gives a relative measure of the residual capacity of the structure. It takes
a value of unity for the intact structure and reduces to zero when the structure turns
into a mechanism. The incremental change, before and after the failure of an element,
gives an indication of the importance of the member to the integrity of the structure.
6-62

Table 6.2: Deterministic failure sequence for the BRITE structure.

ElemNo. 2134 2336 1124 1326 3346 3144 3134 1124


Node No. 304 306 204 206 406 404 301 101
Load Factor 2.98 2.98 3.11 3.11 3.24 3.25 3.42 3.42

DI 0.99 0.82 0.79 0.60 0.59 0.57 0.21 0.13


Elem No. 1326 2134 2336 3336 2326 2124 3134 3336
Node No. 103 201 203 303 203 201 304 306
Load Factor 3.42 3.42 3.42 3.43 3.44 3.44 3.46 3.46
lB I 0.12 0.07 0.04 0.02 .006 .002 .001 .000

The results of the collapse analysis can also be expressed in the form of the following
non-dimensional ultimate strength measures, see Frangopol et al (1991):

Reserve Strength Factor - . - 3.46


Qdes

Design Reserve Factor - Q1m1 -


2.98
Qdes

- o1 - Qimi - 0.20
System Redundancy Factor__________
Q01

= Design environmental loading


Qfmf = Environmental load at first member failure
= Environmental load at structural collapse

The above results are based on only one wave direction (along +ve Y-axis). In a
practical analysis the above procedure should be repeated for a number of wave
directions to determine the lowest collapseL load. From the above liniited result& the
following observanons can be made about the collapse behaviour of this structure:
6-63
* The structure is highly over-designed with respect to the failure of individual
members. It also has a sufficiently high reserve against collapse beyond the
design environmental load. But the effective redundancy of the siructure
beyond first member failure is very low. Although a detailed damage
tolerance assessment is not made it is clear that the structure has poor damage
tolerance properties. This structure clearly illustrates the drawbacks of the
current design practice which is based on strength check of individual
members while all system effects are ignored.

* From the detailed failure sequence it can be seen that, although a total of 16
plastic hinges have formed before the structure became a mechanism, most of
the load carrying capacity is lost when about 7 or 8 plastic hinges develop.
A value of lB =0.3 can be taken as a practical measure of failure for this
structure.

* The collapse mechanism is entirely formed by the failure of members in the


two Y-direction vertical frames while the members in the two X-direction
frames and horizontal diagonal members do not offer significant resistance
against collapse. However, the X-direction frames are expected to be
important to resist against collapse when the wave approaches along X-axis.

* The 6 compressive diagonals in Levels 1/2/3 in the two Y-direction frames are
the weakest and hence the most important members for the integrity of the
structure. However, at least one of the main horizontal brace has to fail before
the structure can reach the collapse state.

* Since the corresponding members in the two Y-frames share equal load they
fail almost simultaneously during the loading process. Moreover, since most
of the diagonal members are dominated by axial behaviour, with equal forces
at either ends, a member does not support any further increase in load once a
plastic hinge forms at one end of the member. This is more close to the
behaviour of a truss member.
6-64
SYSTEM RELIABILITY ANALYSIS

System reliability analysis of the jacket is carried Out using the program RASOS_R.
The program uses most of the information generated in the form of data-base files by
previous runs of other modules such as, structural model, deterministic components
of environmental loading, stiffness matrix, influence matrix etc. Only the minimum
necessary calculations to determine random member strengths, random linear response
and progressive collapse analysis for a prescribed failure sequence are carried out
within RASOS_R thereby increasing the efficiency of reliability analysis.

Probabilistic Modelling:
A list of basic variable types that can be used for probabilistic modelling within
RASOS are given in Appendix-C. Based on a preliminary reliability analysis of
selected members a few variables such as wind speed, inertia coefficient etc. were
found to be relatively less important for member reliability and were considered as
deterministic during subsequent analyses. The final list of basic variable types and
their assumed probability distributions used in system analysis is given in Table 6.3.

Table 6.3: Probability distributions for basic variables.

Variable Distribution Mean COY

Wave Height Gumbel 18.0 m 0.15


Wave Period Lognormal 11.0 sec 0.10
Current Speed Lognormal 0.9 rn/sec 0.20
Marine Growth Lognormal 4.0 mm 0.15
Drag Coefficient Lognormal 0.70 0.20
Load Model Bias Lognormal 1.00 0.15
Yield Stress Lognormal 300 MPa 0.10
in Tension
Yield Stress in Lognormal 300 MPa 0.13
Compression
6-65
The wave height (H), wave period (1') and current speed (Vs) were modelled as
correlated variables with correlation coefficients p=O.6O, p=O.4O and p.=O.3O.
The yield stresses in tension and compression were assumed to be independent
identically distributed for all members in the structure.

Idealization of Member Behaviour:


As mentioned earlier, the member limit-forces in axial tension and bending moments
were calculated based on mean value of yield stress and plastic section properties
while the compressive capacity was determined using elasto-plastic buckling analysis.
Within reliability analysis, these mean-value member capacities were randomized
using random multipliers to reflect the variability in yield stresses. Thus all members
were assigned limit-state type = 1 as per Appendix-B. if a more rigorous assessment
of member capacities is desired or when initial imperfections in members are modelled
as random limit-state type 3 has to be used.

Selection of Failure Elements:


Since the number of potential failure sequences increases substantially with the
number of failure elements, every effort is made to reduce the number of failure
elements without unduly sacrificing the accuracy of the analysis.

Based on the results of linear analysis, progressive collapse analysis and preliminary
reliability analysis of isolated members (Level-O) analysis, 32 members (from out of
a total of 156 members), from the two Y-direction frames between levels 1 to 4 were
selected for inclusion in system analysis. All other members in the X-direction
vertical frames, all horizontal diagonals and all members above the mean sea level
were considered stochastically unimportant.

Although RASOS_R allows consideration of plastic hinges at both ends of a member,


based on the results of deterministic collapse analysis and considering the dominance
of axial behaviour of most diagonal braces, only one plastic hinge per member was
used during reliability analysis. This will reduce the number of failure elements by
half and a substantial reduction in the number of failure sequences to be analyzed.
6-66

Definition of System Failure:

A complete failure of the structure corresponds to the formation of a collapse


mechanism as indicated by a zero value of the determinant of the active influence
matrix, (l B =0). However, a large deflection at the deck level, which makes the
platform unsuitable for continued operation, or a substantial loss in load carrying
capacity may be used as more practical measures of failure. From the deterministic
collapse analysis it was seen that the BRITE structure looses nearly 95% of its load
carrying capacity when about 7 or 8 plastic hinges are formed and the value of B
falls to about 0.3, although a total of 16 plastic hinges are required to turn the
structure into a mechanism. For reliability analysis, therefore, system failure event
was defined as lB ^0.3. Initial trials showed that there is very little increase in
reliability beyond this level. Moreover, branches from a partial failure path at level
8 or above will often be highly correlated and contribute very little to system
reliability. This failure criterion reduced the computational effort substantially since
the length of complete failure paths and the number of possible paths were reduced.

Choice of a Truncation Criterion:

The failure-tree enumeration approach proposed in section 6.6.3 requires a truncation


criterion, for discarding non-dominant sequences, to be specified before the fault-tree
analysis can begin. It was argued that a good initial estimate of system reliability can
be obtained by directly evaluating the probability of the failure sequence obtained by
a deterministic collapse analysis. Following the sequence reported in Table 6.2, as
closely as possible, and with the failure criterion defined above, a reliability index of
was obtained for the BRITE structure. Based on this it was arbitrarily
decided that a partial failure path with (1343*)>1 .05 does not contribute significantly
to system failure probability and hence further branchings from this path can be
stopped. During failure-tree enumeration if a complete failure path with 1 3 lower thai
2.92 is found, the value of 3* can be reset to this lower value.
6-67
Overview of Failure-Tree Enumeration:

Failure-tree enumeration is carried out using 32 important failure elements selected as


discussed earlier. Although RASOS_R offers an automatic failure-tree enumeration
the interactive option was used for the analysis. In lieu of an exhaustive enumeration,
which could be very time-consuming, an attempt was made to obtain close bounds on
system reliability with minimum computational effort.

Starting from the intact state of the structure, reliability analysis was carried out for
each of the 32 members and the results of reliability indices for some of the important
members are shown in Fig.6.7. The notation used for the failure-tree is explained in
Fig.6.6. From these results it can be seen that only 6 of the 32 members have
significantly lower reliabilities than the remainder while branching from the remaining
elements can be stopped based on the chosen truncation criterion. The union over all
these failure events gives a reliability index for first member failure (system reliability
at Level 1) as =2.3O. This also gives a crude estimate of the lower bound on
system reliability. The 6 most dominant members happen to be the compressive
diagonals in the two Y-direction vertical frames. Corresponding members in the two
frames (viz. 213412336, 1124/1326 and 3144/3346) are seen to have nearly equal
reliabilities.

Since member 2134 has the lowest reliability it was selected first for further branching
to Level 2. For this purpose, safety margins of all surviving (32-1) elements are
calculated with values of the basic variables corresponding to the (joint) beta-point of
member 2134 and are arranged in increasing order of their magnitudes. Since a
truncation criterion is available in advance non-dominant branches can be discarded
at every stage of the analysis. The branches obtained in this way from members 2134
and 1124 are shown in Fig.6.7. It was seen that branching from element 2336 results
in partial failure paths with identical reliabilities to those obtained from member 2134.
All branches from elements 3144 and 3346 result in partial failure paths with high
reliabilities which can be discarded by the truncation criterion.
6-68
The analysis was similarly continued to Level 3, the results of which for only few
paths are shown in the figure. At this stage it became clear that only branches starting
from elements 2134 and 2336 would be dominant. For the subsequent analysis,
therefore, an attempt was made to first trace a set of dominant complete failure paths
starting from member 2134 in order to obtain an upper bound on system reliability
quickly. These dominant failure paths for the structure are shown in Fig.6.8. A
similar set of dominant mechanisms can be obtained by starting the enumeration from
element 2336 and these are shown in Fig.6.9. An upper bound on system reliability
of 3=2.86 is obtained by considering all the identified complete failure paths.

Although, nearly 30 most dominant mechanisms have been identified for the structure,
it is clear that many more dominant mechanisms can be obtained by different
arrangement of failure elements. For example failure sequences starting from elements
1124 and 1326 are also expected to be dominant. However, from the sequences traced
so far, it can be seen that the 6 six compressive diagonal members 1124, 1326, 2134,
2336, 3144, 3346 are present in many of these dominant failure paths. An attempt
was then made to trace alternative failure sequences which did not involve these
dominant elements. However, it was found that any partial failure path which did not
involve at least 5 of these 6 elements will lead to a non-dominant complete failure
path as can be seen from the sequence (2134+1124+3144+2936+2734) shown near the
bottom of Fig.6.8. Any additional failure sequences containing the same 6 elements
will all be highly correlated with those already identified. Not much improvement in
the upper bound can be obtained by identifying additional failure paths. Instead it is
more advantageous at this stage to improve the lower bound on system reliability. A
lower bound of LB=244 is obtained by considering all dominant partial failure paths
up to Level 3. By extending only the most dominant of these sequences up to Level 5
the lower bound can be increased to a value of I3LB=260 The final bounds on system
reliability from the present analysis can be given as:
2.60 2.86
These bounds are believed to be sufficiently close for practical purposes and the
analysis is stopped at this stage. The true value of is expected to be close to the
upper bound.
6-69
Discussion of Results:

From the results of the reliability analysis at Level 1 it can be seen that only 6 of the
32 members considered have significantly lower reliabilities compared to others.
These members are the compressive diagonals (1124, 1326, 2134, 2336, 3134, 3346)
in the two Y-direction vertical frames along the wave propagation direction.

The set of complete failure paths identified for the structure are all made up of the 6
most dominant members. Any failure sequence which does not contain at least 5 of
the 6 dominant members has a relatively high reliability index and consequently does
not contribute significantly to the system failure probability.

The failure sequence obtained from the deterministic collapse analysis has a reliability
index of 2.92. Due to convergence problems the exact sequence as given in Table 6.2
could not be analysed. However, a sequence which is very close to this, containing
essentially the same elements, was analyzed with the chosen criterion for system
failure and this is marked "dcs" in Fig.6.8. Comparing this with other identified
failure paths it can be seen that, although the deterministic collapse analysis gives one
of the most dominant failure paths there are few other failure paths which are equally
dominant. For example, failure of member 3336 in place of 3134 or non-failure of
member 3144 or a different order of failure of members will also give also give same
reliability as the "dcs" sequence. However, these sequences are not structurally very
different from each other.

The results of sensitivity factors for one of the failure sequences are shown in
Table 6.4. From this it can be seen that the reliability index is highly sensitive to
loading variables while the resistance variables show very little influence. Of the
loading variables wave height is the most dominant variable followed by wave period
and drag coefficient. Of the resistance variables the reliability index is seen to be
more sensitive to the uncertainty in yield stress values of members 2134, 2336 and
3134 which also indicate the relative importance of these components to the failure
of the structure through this failure path.
6-70

Table 6.4: Sensitivity factors for a complete failure path

Variable H T V tmg Cd

a -0.776 -0.374 -0.038 -0.259 -0.335 -0.258


Variable F234 F1326 F36 F3144 F1124 F3134
a 0.084 0.012 0.041 0.022 0.010 0.040

Table 6.5: Summary of results for the extreme wave condition

Reliability Index for: Value



Member most likely to fail first Im1fm 2.43

First failure of any member IL1 2.30

Most likely complete failure path m1fp 2.92
System failure through any path
I3sysUB 2.86

The dominance of loading variables, which are common to all elements, introduces
high correlations between failure events of different components within a failure path
and between failure paths. The correlation coefficients between individual elements
were found to be in the order of 0.92 while the correlations between the complete
failure paths were as high as 0.98.

The final results of the reliability analysis are summarized in Table 6.5. From this it
can be seen that the difference between the system reliability index and failure of any
one member (Level 1 reliability index), which represents the so called "system effect",
is very small. The results can also be expressed in the form of probabilistic
redundancy measures as, see Frangopol et al (1991)
6-71

L1 -
Complexity Factor - 0.95
Imlfm

tmlfm
Net System Factor - _____ - 0.85

Redundancy Factor - - L1
________ - 0.20

The complexity factor expresses the effect of number of elements, relative dominance
of elements and correlation between components at Level 1. A value closer to unity,
as seen above, implies high correlation between components and relative dominance
of only a few members. Also the high value of net system factor implies that the
system effects are not prominent for this structure. Similarly a low value of
redundancy factor implies that the structure has a high probability of reaching final
failure given the initial failure of any one of its primary members.
5U U 5U
I-

I I .3
I
.3
C
I 0
I. 0
C.)
0
0

U
S
0
H .b
I,
•0 • .0

c4Z
01-b
•0 I
U
V

H
C
01

• —
.0 E
U C

0
U
-I
.3
0)01
U

3.52
3744

3.27 2.59
3946 2734 2336
2.44 2.59 2.82
2338 1124 3346

2.43 2.59 3.08


2134 1328 2936

3.31 2.72 3.09


2734 3348 2936

3.31 2.72 3.25


2938 3144 2734

2.72 3.28
3346 2938

2.72 2.59 2.50


3144 2134 2338

2.59 2.83 3.29


1124 3144 2734

3.32 3.24
1328 2734 2936

5.23
3134

2.30 p 2.35 2.44


L2

Fig.6.7: BR1TE structure: Failure sequences up to level 3


c1
I-

c.J
0
cJ

c.J
0

3.16 2.92 3.08


\—f33e 13134

3.08
3330

3.04
2.58 2.83 I 2.83 2.83 2.94

132\-(3144' 1-(2 13 41-ti 124 2734
3.04
2.82 3.04
2.44 3346
2.92 2326J /
?330 / 3.04
3336
I 11926
2.58 2.58 2.70 2.70 2.94 2.94
1 34 3345 132 0 2936 3144

2.82
3144

2.92

Fig.6.9: Failure sequences starting from member 2336


6-77

Example 2: Reliability of the Jacket Structure Under Fatigue Deterioration

The analysis in the previous example entirely ignored the effect of fatigue damage on
the reliability of the structure. In this example the influence of fatigue is investigated
by treating it as a deterioration in the strength of tubular joints under plastic collapse
and fracture limit-states.

A quasi-static spectral response analysis of the structure is first carned out to


determine the long-term distribution of hot-spot stress-ranges at various tubular joints
in the structure. The long-term wave climate is represented by 6 sea-states and two
directions of wave approach are considered. The stress concentration factors are
calculated using Wordsworth/Smedley formulae, see U.E.G (1985). The probability
density function of stress-ranges is computed using the short-term and long-term
models developed in chapter 3. The module RASOS_H is used for this purpose.

Next, a deterministic assessment of a number of joints is carried out to determine


criticality of the joints under the limit-states of fatigue, plastic collapse and
fatigue-fracture. The program module RASOS...F is used for this analysis. Based on
these studies a total of 10 tubular joints are identified for use in a system reliability
analysis. These joints happen to be the ends of tubular X-braces in the two vertical
frames parallel to Y-axis. The wave approach directions considered for fatigue
loading analysis are symmetrical about the Y-axis.

Reliability analysis of the above 10 joints is first carned out for the limit-states of
plastic collapse and fatigue-fracture using the methodology of chapter 5. The
probabilistic modelling of variables is also similar to that used in the examples of
chapter 5. The program module RASOS_R is used for this analysis. The results of
the analysis allowing for fatigue exposure over 20 years are given in Table 6.6 below.
6-78
Table 6.6: Results of reliability analysis for tubular joints under fatigue

Joint Number Plastic Collapse Fatigue-Fracture

1124/204 3.121 3.104


1724/204 4.193 4.238
2134/304 2.709 2.624
2734/304 3.8 10 4.250
2336/306 2.713 3.519
2936/306 3.811 4.238
3144/301 5.009 3.810
3744/307 6.056 2.746
3346/303 5.003 2.708
3946/309 6.047 4.204

From this table it can be seen that the reliabilities for tubular joints are relatively
much higher than the tubular members. Tubular joints loaded under axial compression
have lower reliabilities than those loaded under tension. For the fracture limit-state
all joints are considered to be in as-welded condition with mean tensile residual
stresses of 300 MPa. Out of the 10 joints only 2 joints, namely 3744/307 and
3346/303, showed influence of fatigue on the limit-state of fracture. The joint
reliability under the limit-state of plastic collapse did not show any influence of
fatigue for any of the joints.

When these 10 joints are combined with the 32 tubular members considered in the
previous example, it is to be expected that the system reliability of the structure will
not change. Since the reliabilities of all the tubular joints are much higher than the
dominant tubular members, sequences involving the tubular joints result in high
reliabilities and consequently do not contribute to the system failure probability.
Moreover, since only two joints show a small influence of fatigue deterioration, the
system reliability of the structure does not show any variation with service exposure.
6-79

Thus, at least for this example, the system reliability of the structure is entirely
governed by the tubular member failure failures for which an extreme wave reliability
analysis is adequate.

However, when the limit-state of fatigue (through-thickness cracking) is considered,


as investigated in example 1 of chapter 7, it is seen that the reliability of few tubular
joints are significantly lower than the tubular member limit-states. Thus, at least for
this example, the approach of accounting for fatigue effect in the form of deterioration
in joint strength under plastic collapse or fracture modes does not seem to capture the
influence of fatigue correctly. However, these observations cannot be easily
generalized as the fatigue-fracture limit-state is influenced by many factors, such as
fracture toughness, level of residual stresses, direction of extreme wave approach with
respect to the dominant direction of fatigue waves, etc as discussed in chapter 5.
Explicit consideration of the fatigue limit-state requires time-dependent system
reliability formulation and this is investigated in the next chapter.

6.8 CONCLUDING REMARKS

The current state-of-the art on reliability of offshore platforms has been reviewed in
terms of the methods used for modelling element post-limit behaviour, structural
collapse analysis, treatment of environmental loading, reliability formulation for a
failure sequence and enumeration of failure-tree. Although the available methods are
fairly well developed for the analysis of small structures with idealized material/
member behaviour models, they are not yet suitable for more realistic analysis of
practical offshore structures. The following main limitations have been observed:

(a) A simplified stochastic representation is used for the treatment of uncertainty


in environmental loading. Often nodal wave loads or the base shear are
modelled as random variables the variability of which are related to the
variability in wave height alone. The problem in including all environmental
6-80
variables explicitly is the computational cost involved in including full
hydrodynamic force calculations within iterative reliability analysis.

(b) For structural collapse analysis required in determining system failure events,
either plastic mechanism based methods or successive elastic analyses with
member replacement are used. This limits the methods to the treatment of
strain-independent member post-limit behaviour such as elastic-ideal plastic or
elastic, semi-brittle models. More realistic softening/hardening behaviour
cannot be treated using these methods.

(c) The reliability formulation for a failure sequence is based on the so called
crude FORM approach in which the elemental failure surfaces are linearized
in their respective beta-points which in some cases can occur in physically
inadmissible probability space. This precludes the use of standard non-linear
structural analysis methods which can treat more realistic member behaviour.

(d) The methods used for the identification of dominant failure paths are
undoubtedly more rigorous and accurate but they turn out to be very
impracticable for the analysis of full-scale offshore structures.

(e) The tubular joint behaviour and fatigue effects are not accounted for in an
extreme wave reliability analysis mainly for the limitation imposed by the
member replacement approach which allows only one failure mode for each
structural member.

A new system reliability analysis procedure is presented in this chapter which helps
to overcome the above limitations in an unique way as summarized below:

(a) A response surface approach is used for the calculation of extreme


environmental loading. This allows explicit modelling of uncertainties in wave
height, wave period, water depth, current speed, wind speed, marine growth
6-81
and drag/Inertia coefficients. The random linear response due to environmental
loading can be calculated very efficiently during iterative reliability analysis.

(b) A non-linear collapse analysis based on a virtual distortion method is used


which simulates elemental failures in terms of a set of virtual distortions to
determine the modified stress/strain state of the structure without the need for
re-formulation of the stiffness matrix. The method is very efficient compared
to the successive elastic analyses approach or other recently developed
incremental non-linear analysis procedures. This allows the treatment of more
realistic member post-limit behaviour including softening/hardening effects.

(c) The reliability formulation for a failure sequence is based on the joint
beta-point concept in which the elemental failure surfaces are linearized at the
minimum point of the intersection failure domain. The search for the joint
beta-point, which is always in a physically admissible domain, is carried Out
using a multi-constrained optimization technique. This allows the use of
powerful non-linear analysis methods, such as the virtual distortion method,
without straying into physically inadmissible regions of the probability space
and the accuracy of the computed reliability is also significantly improved.

(d) A new method of failure-tree enumeration has been developed which uses
selective enumeration and preferential branching to identify failure sequences
in the order of their importance and computes bounds on the system reliability
to within user specified limits. The method is to some extent approximate but
is highly efficient compared to the previously available rigorous methods and
makes it suitable for the analysis of full-scale jacket structures.

(e) The use of the modified virtual distortion method of non-linear analysis makes
it possible for the inclusion of tubular joint limit-states of plastic collapse and
ductile fracture as explicit failure elements in a system reliability analysis.
The effect of fatigue can be included in a simplified form as a deterioration in
6-82
joint strength as a function of crack size. A probabilistic fracture mechanics
model is used for quantifying crack growth with time.

With these refinements the proposed reliability methodology represents a significant


improvement in the state-of-the art and makes a practical system reliability analysis
of full-scale jacket type offshore structures possible. The proposed methodology has
been implemented in the computer software RASOS developed under the CEC BRITE
project in association with other researchers. The methodology and the software has
been demonstrated on a shallow water jacket structure considering only a limited
number of failure elements. More detailed case studies using full-scale analysis need
to be performed to evaluate the potential of the method for design and assessment of
offshore structures.

The proposed methodology still has certain limitations in common with all presently
available methods. The system reliability formulation is essentially time-independent
and uses only random variable models for description of uncertainty in basic variables.
Random process models for loading and dynamics of structural response (inertial
effects) cannot be treated. Consequently the applicability of the method is restricted
to jacket structures in shallow to medium water depths and cannot be used for
semi-submersibles and tension leg platforms. The method is simplistic in the
treatment of fatigue effects which is essentially a time-dependent phenomenon and
consequently the fatigue limit-state defined by through thickness crack formation,
which is one of the most dominant modes of actual joint failure cannot be explicitly
treated.
Chapter 7
TIME-DEPENDENT SYSTEM RELIABILITY ANALYSIS

7.1 GENERAL

The system reliability method given in chapter 6 is essentially a time-independent


formulation in which all elemental failures, leading to system failure, are considered
to occur under a single load event. The applicability of the method has been
demonstrated for the reliability analysis of jacket structures under a life-time extreme
wave event. A simplified approach for including fatigue effects in the form of
deterioration in joint strength was also presented. In this case, however, only the
annual probability of failure could be evaluated for the structure. Moreover, the
limit-state of fatigue failure (through-thickness crack formation) of joints cannot be
considered in such an approach.

In jacket-type offshore structures fatigue failure has been found to be the dominant
limit-state for most tubular joints. In a redundant structure fatigue failure of one of
the joints will cause increased cyclic stresses in the remaining joints. If this joint
failure is not detected and repaired the increased cyclic stresses may lead to a second
joint failure or a series of joint failures leading to the collapse of the structure. The
sequence of fatigue failures invariably occur at different points in time. Alternatively,
following fatigue failure of one joint and accelerated fatigue crack growth in others,
additional joints (deteriorated) may fail by fracture during moderate sea conditions
(seasonal storm extremes) leading eventually to the collapse of the weakened structure
in an extreme storm event. In the absence of in-service inspection such a failure
scenario is considered to be more realistic for jacket structures than the scenarios
previously described. Reliability methods to deal with such scenarios is presented in
this chapter. These could be used as rational tools in the planning of in-service
inspection and maintenance programmes.
7-2
A few methods for the reliability evaluation of a sequence of fatigue failures have
been reported in the literature and these are reviewed in section 7.2 to identify their
limitations. A new and more general formulation for time-dependent reliability
analysis is developed in section 7.3. Application of this formulation to the analysis
of a sequence of pure fatigue failures or a combination of fatigue and fracture failures
of tubular joints is presented in sections 7.4 and 7.5. Finally, the methodology is
extended in section 7.6 to the treatment of combination of time-dependent and
time-independent failure sequences. The method is demonstrated through examples.

7.2 REVIEW OF AVAILABLE TIME-DEPENDENT RELIABILITY METHODS

In a time-dependent problem it is necessary to define the time-scale of the reliability


analysis precisely. In the context of offshore jacket platforms three time domains,
defined primarily on the loading regimes, are of interest. The ability of the structure
to survive under these loading regimes can be studied separately.

(a) The survival of the structure as the life-time extreme wave passes through the
structure. The time duration is typically 10-20 seconds.

(b) The ability of the structure to survive in a life-time extreme storm sea-state.
The time duration in this case is typically 5-8 hours.

(c) The ability of the structure to withstand the continuous action of wave, wind
and current forces throughout the life time of the structure.

Conventional design practice treats these loading conditions differently and reliability
analysis methods can similarly be formulated separately for these cases. In the first
case the effect of time arises mainly due to the phase difference between the different
partsof the structure and-thepassing wave. It is well known That the forces in
individual members vary as the wave passes through the structure. However, to avoid
repetitive structural analysis the design practice has been to find the position of the
wave crest which maximises the horizontal base shear for that direction and then
7-3
determine member internal forces for that wave position. The reliability analysis
methods also adopt the same approach. Some studies have been performed to assess
the probabilistic significance of this phase effect, see Turner and Baker (1988) and
Hoim et al (1988). It has been found that the reliabilities of individual members vary
with wave position resulting in small changes in the relative importance of different
failure paths. The overall system reliability also changes but was found to be lowest
for the wave position for which the base shear is maximum. Consequently it is
considered adequate to assess the safety of the structure for the critical wave position.
The time-independent system reliability methods discussed in the previous chapter can
therefore be used to deal with this case. This discrete wave approach is considered
to be satisfactory for most shallow water jacket structures.

For compliant structures and deep water jacket structures for which dynamic effects
are important a discrete wave approach is not satisfactory. In such cases, (loading
regime-b) the conventional design practice has been to assess the safety of the
structure under a life-time extreme storm sea-state using a spectral dynamic analysis
procedure (more precisely a design storm of 100-year return period is used). Such a
storm based approach can also be used in a reliability assessment. Baker and
Ramachandran (1981) have used a reliability analysis procedure based on spectral
dynamic response analysis to determine reliability of individual members of a jacket
structure. However, the spectral response analysis methods are valid only within the
linear-elastic regime and consequently progressive collapse and system reliability
cannot be assessed using this approach. A rigorous uncertainty modelling of
time-varying loading will also require a random process formulation of reliability
which is still under development. Guers et al (1988) have developed a system
reliability method for simple brittle redundant structures using an uperossing
formulation of a stationary, Gaussian scalar load process.

When dynamic effects are significant and/or more than one dominant time-varying
loads are present it is difficult to determine a unique extreme loading condition for
assessing the safety of the structure. In such cases Inglis et al (1985) have advocated
the use of a semi-probabilistic long-term method to determine the extreme response
7-4
statistics of the structure considering the contributions of a number of sea-states which
may occur over the life of the structure (loading regime-c). A similar long-term
reliability analysis method has also been attempted by Bjerager et al (1988) using a
combination of random process and random variable reliability formulations.

Another long-term consideration of interest is the cumulative fatigue damage. Since


fatigue is a cumulative and time-dependent phenomenon elemental fatigue failures can
occur at different points in time. In addition deterioration in joint strength as a result
of fatigue may lead to fracture failure of joints at different points in time under the
action of multiple load peaks. The system reliability methods in the following address
this aspect of the problem considering a time-scale extending over the life-time of the
structure. The short-term effects of time-varying loading regimes (a) and (b) on
system reliability are ignored.

Studies on system reliability of jacket structures under fatigue conditions are relatively
scarce. First of such studies is due to Martindale and Wirsching (1983). The method
is based on an S-N approach for fatigue damage and a simple lognormal format (see
chapter 5 for details) for reliability evaluation of isolated joints. A sequence of pure
fatigue failures is considered in which the time to system failure is evaluated as the
sum of time intervals between successive failures. The developed methodology is
applicable for parallel member systems with joints of equal or unequal strengths but
sharing equal loads. The probability distribution of time to system failure has been
derived for 2, 4 and 8 member parallel systems. Even for such simple cases the
problem was so involved that Monte-Carlo simulation had to be used. The effect of
correlation in fatigue lives between joints as a result of common cyclic loading has
not been addressed.

Using the same lognormal format for joint reliability Stahl and Geyer (1984)
developed an analytical distribution for time to system failure. The formulation
however is limited to equal strength and equal load sharing parallel system (Daniel's
system). The effect of correlation in fatigue lives has been considered by splitting the
fatigue life into a dependent component and an independent component. The system
7-5
reliability formulation is based on the independent component of fatigue lives of
various joints for which the probability of failure is calculated using multi-dimensional
(equal to the number of components) numerical integration. The independent
component is finally convoluted around the dependent basic variables to obtain overall
system probability of failure.

The above approach has recently been extended to the case of unequal strength and
unequal load sharing systems by Kumar and Karsan (1990). The essence of this
method is to express the conditional probability distribution of time to failure of a
member following failure of other members in terms of its base (unconditional)
probability distribution. The member failure in this study has been defined as the
failure of either of the two joints at the ends of a member. An identical method has
been used by Karsan and Kumar (1990) in which a system failure is defined,
pragmatically, in terms of a sequence of joint failures and not member failures. A
branch and bound type of procedure has also been used to identify stochastically
dominant failure sequences. In both these methods, however, numerical integration
has to be used to compute system failure probabilities.

The above methods of fatigue system reliability, however, have the following
fundamental limitations:

(a) The methods are based on S-N approach which provides an explicit expression
for time to fatigue failure but cannot be used with the fracture mechanics
approach. The applicability of the methods for inspection planning and
reliability updating is therefore limited.

(b) All the fatigue basic variables have to be assumed to follow lognormal
distribution so that a lognormal model for time to failure is easily obtained.
More general probabilistic descriptions cannot be used. Similarly a general
correlation structure between variables is not possible.
7-6

(c) The latter methods use multi-dimensional numerical integration to evaluate the
probability distribution for the independent component of time to system
failure. The application of these methods is thereby severely restricted if
failure sequences of more than 4-5 components have to be analyzed.

A new and more general formulation for time-dependent system reliability analysis to
deal with long-term fatigue effects is developed in the following.

7.3 TIME-DEPENDENT SYSTEM RELIABILITY FORMULATION

A time-dependent system reliability formulation for a jacket structure is developed in


the following in which elemental failures can occur at different points in time during
the service life of the structure. The method is general in terms of probabilistic
description and correlation structure for basic variables. The methodology is based
on the use of efficient FORM/SORM techniques for probability computations.

7.3.1 Time-Dependent Prescribed Failure Sequence:

Consider a jacket structure undergoing deterioration due to fatigue at its tubular joints.
Consider a failure sequence k: l,2,...,j,...,q in which the tubular joints fail at random
time instants t1 , leading to the collapse of the structure. Let t be the
intended service life of the structure. Obviously, the system is considered to have
failed if all joints fail within the service life of the structure thereby rendering it unfit
to fulfil its intended purpose. Then the failure sequence can be expressed as an
intersection of elemental failure events and its probability evaluated as

Pf.k - P [(t 1^r)fl(t2^Ofl. .fl(t^t)ft . •fl(tq^t)]


* (7.1)
7-7
In general the failure of one tubular joint will increase the stress levels in the
surviving joints which will increase the rate of damage thereby reducing their
remaining lives. The failure times i in the above are therefore conditional on the
failure of the previous joints and the time instants at which they fail.

The time instants 13 are generally not known but these can be easily expressed in terms
of the base lives of joints. It is convenient to introduce the following notation:

Si base stress in joint j corresponding to intact state of the structure


base damage rate of joint j corresponding to stress level S
T base time to failure of joint j undergoing deterioration at rate
1)1
actual time instant at which joint j fails, and
S the stress level in joint j during the iII) interval after (i-i) joint failures
the damage rate corresponding to S3 after (i-i) joint failures
(i)
1
the 'damage ratio' giving the ratio of the base damage rate in joint j
to the damage rate during the ith time interval,
the remaining life of joint j after 1 ,2,...,(i- 1) joints have failed in that
sequence. Obviously t1=t'.

The base fatigue life of a joint T can be determined using an appropriate theory, for
example using an S-N or a fracture mechanics model. Assuming linear damage
accumulation (Miner's hypothesis) life of a joint will be inversely proportional to the
damage rate. If the change in damage rates following successive joint failures can be
determined, the remaining life of a joint can be related to its base life.

The change in remaining life of a joint after each successive joint failures is shown
schematically in Fig.7.1 for a failure sequence l—*2--3-4. For the sake of simplicity
the stress level in a joint is shown to always increase after the failure of some other
joint. For joint 1, i.e. the first joint to fail in the sequence, the time-to-failure is
obviously given by

- - T1 (7.2)
7-8

(1)

(2)

(3)

4)

Fig.7.1 Illustration of change in remaining lives after successive failures-


7-9
After the failure of the first joint, the stress levels in the remaining joints (2,3 and 4
in the figure) change and correspondingly their remaining lives also change. The
second joint, for example, is stressed at its base stress-range S 2 for the duration t 1 and
at the modified level S 2 over its remaining life. Referring to Fig.7.1, the remaining
life of joint 2 after the failure of the first joint and hence the time to second failure
can be calculated as


(2) - [T2-t1] - [T2-t1]y
(7.3)

t2 -

Next, consider the failure of joint 3. Its stress level changes once after the failure of
the first joint and again after the failure of the second joint. The remaining life of
joint 3 after the failure of the second joint thus also depends on the change in its
damage rate after the first joint failure. The time-to-failure of the third joint can be
determined by calculating its remaining life after successive joint failures as

(1)
r3 -T3

(1)
- - [T3-:1]y2
3
(7.4)
(2) (3)
(3) (2) d3
- [t - (t2 -r1 )] [r3 - (t2 -t1 )] -
(2)
'13

(3 - 12+t3

The above expressions for remaining life can be generalized to higher levels as

(i)
(1) '1
_________
- [t (-" - (r t_2)] (s-i) (7.5)
Ti
7-10
with t0 =O and f=i by definition. The actual time instants tj for successive joint
failures can now be determined as

()
tI - 'C i - P
it
1) (2) -
, + +t
(7.6)

The discussion in the above has been presented in the context of fatigue failure of
tubular joints for simplicity and ease of understanding. The above formulation can in
fact be used for any time-dependent problem if appropriate expressions can be derived
for failure times T and factors %(1) to quantify the effect of previous elemental
failures. However, it is important to note that the elemental failures have to be brittle
so that once an element fails it will remain in a failed state through the remaining
service life of the structure. If an element fails in a ductile mode under a load peak
it will revert back to a safe state once the load event has elapsed and a time-dependent
failure sequence cannot be formulated in such a case. In general terms the elemental
failure events can be expressed in terms of conditional safety margins, which for the
Jth
elemental failure can be given as

M' - (1_1+(,i)) - (7.7)

and the probability of the failure sequence can be evaluated in the usual way as

Pf,k - P[(Mi^0)fl(M2^0)fl...fl(Mj^0)fl...fl(Mq'^0)] (7.8)

In the context of offshore jacket structures the only forms of time-dependent brittle
failure modes of interest are the fatigue failure of tubular joints and fracture of joints
under a storm peak following fatigue deterioration over an arbitrary period of time.
Expressions for conditional safety margins for fatigue failure events are developed in
section 7.4 while for the fatigue-fracture limit-state are derived in section 7.5. The
base fatigue hves and actual times-to-failure are expressed in terms of basic random
variables described by arbitrary probability distributions and correlation structure.
7-11
7.3.2 Time-Dependent Failure-Tree Enumeration:

The failure-tree enumeration method developed in chapter 6 (section 6.6.3) is


relatively independent of failure modes for individual failure elements and the manner
in which a failure sequence is formed. In particular, the truncation criterion, selective
enumeration, branching criterion and the basis for termination of the analysis can be
equally applied for the case of a time-dependent failure-tree. However, the failure
elements should now include only fatigue and fatigue-fracture limit-states for tubular
joints. The selection of critical joints and critical limit-states can be made based on
preliminary deterministic analysis and reliability analysis of isolated joints from which
joints can be categorised as fatigue-critical and/or fracture-critical. A deterministic
fatigue failure sequence, leading to the collapse of the structure, can also be obtained
by subjecting the structure to fatigue cycling over an unlimited period of time and
updating the hot-spot stress distributions after each joint failure. A reliability analysis
can then be carried out for this single failure sequence to obtain the probability p
of the most dominant failure mode for use in the truncation criterion.

7.3.3 Calculation of Failure Probabilities:

If the uncertainty in all basic variables is described in terms of random variable


models, the methods developed for the time-independent case can be used without any
modifications for the calculation of system probability of failure from a
time-dependent failure-tree. For a prescribed failure sequence, represented as an
intersection event as in Eq.7.8, the joint beta-point can be found using a
multi-constrained optimization technique (NLPQL) as discussed in section 6.5.3. This
failure sequence can then be approximated by an equivalent hyperplane. This is
repeated for each partial and complete failure sequences identified during failure-tree
enumeration. The upper and lower bounds on the system probability of failure can
be calculated from a union over all incomplete and complete failure sequences,
respectively, of the time-dependent failure-tree as discussed in section 6.6.4.
7-12

7.4 ANALYSIS OF A SEQUENCE OF FATIGUE FAILURES

The time-dependent reliability analysis formulation developed above can be used to


analyze the failure of a jacket structure by a sequence of fatigue failure of tubular
joints. Since fatigue failures are brittle in nature there will in general be large
increases in joint stresses due to load shedding from failed members which will
significantly reduce the remaining fatigue lives of surviving tubular joints. If such
joint failures are not detected and repaired they can be potentially hazardous for the
platform. Conversely, a rational planning of in-service inspections and repair
decisions should be based on system considerations quantifying the true risk of failure
of the jacket. In the following safety margin expressions are formulated for a
sequence of fatigue failures. In particular this will involve deriving expressions for
base fatigue lives T1 in terms of basic random variables and damage ratios after
successive joint failures. Both S-N and fracture mechanics approaches are considered
for quantifying fatigue resistance but first an approximate method for deriving the
damage ratio is presented which is common to both the approaches.

7.4.1 Approximate Method for Deriving the Damage Ratio y

Strictly speaking, after each joint failure, the stress levels and stress spectra for each
of the surviving joints due to each of the sea-states should be recomputed and the
damage rates and remaining lives determined using the modified spectra. However,
this can be highly expensive considering the complex nature of spectral calculations
and the large number of failure sequences that have to be analyzed. If the structural
response can be assumed to remain quasi-static, even after several joint failures, the
modified damage rates for the joints can be determined approximately as below.

A base fatigue damage rate for a joint, in a particular sea-state, corresponding to the
intact state of the structure can be defined as

d - 1E[S m ] (7.9)
7-13
where is the average frequency of stress peaks, S is the stress-range and m is the
material constant (exponent) in the S-N or Paris law. A number of methods have been
presented in chapter 3 for determining the mtI moment E[Sm] based on the probability
density function used for modelling hot-spot stress-range. In the case of a
narrow-banded Gaussian stress response process, the damage rate can be given as


D (G 2/Y)mf[m ^2 ) (7.10)
d-

where a is the r.m.s. value of the stress process. For wide-banded density function
models the mth moment will, in addition, be a function of the spectral band-width.

The fatigue loading is typically calculated using an average annual scatter diagram
which is assumed to represent long-term wave climate over the entire service life of
the structure. Moreover, when a tubular joint fails by fatigue the corresponding
member is still retained for the calculation of wave forces. Because of this the
transfer functions from sea surface elevation spectrum to the spectra of nodal wave
loads remain unchanged. The failure of a joint therefore affects only the transfer
function from nodal wave loads to member internal forces which can be written as

- [B11] (HQ (0)). (7.11)

where H(.) are transfer functions of nodal loads, H(.) are transfer functions of
member stress resultants and B is a matrix of influence coefficients. Note that for
quasi-static response the influence coefficients are independent of frequency and the
stress resultants, and hence hot-spot stresses, are linearly related to nodal loads.

If the response of the platform is assumed to remain quasi-static, even after failure of
one or more joints, the average rate of peaks and the spectral bandwidth for a joint
remain unchanged. Consequently, the damage ratio between the base rate and
modified rate after (j-1) number of joint failures becomes
7-14

y(J) - (7.12)

where a is the r.m.s. value of the modified stress process for the j r" time interval.

The influence coefficients b in Eq.7.1 1 change after failure of a joint and have to be
recomputed from the modified stiffness matrix. However, these are independent of
frequency and load magnitude so that, for a given wave direction and load pattern, the
fatigue damage ratio can be wiitten as

m I \m
a I Is I (7.13)
1J =J
where S and S are stress levels in the joint, corresponding to the intact state and
(j1)th damaged state of the structure, computed from a deterministic static analysis

using a single wave of arbitrary wave height approaching from the same direction.

The changes in stress levels following failure of members/joints can be computed very
efficiently using the virtual distortion method of non-linear analysis based on a
deterministic static approach. This greatly simplifies the arduous task of recomputing
spectral characteristics and damage rates following fatigue failure of tubular joints.

7.4.2 Analysis Using an S-N Approach:

An expression for base fatigue life of a tubular joint can be derived easily using an
S-N approach. Published S-N curves are normally based on number of cycles to
through-thic1aess cracking. However, it is observed that a joint can sustain a
significant number of stress cycles beyond this stage dunng which a through crack
propagates around the welded intersection. In a system reliability analysis it is
important to account for this reserve in fatigue strength and an expression for total
fatigue life of a joint at its base stress level can be given as
7-15

T 'R"FL


- (7.14)

'FL - (J)lBg(B:qrE[Sm]

where rli^1 is a multiplication factor on life to through-thickness cracking to obtain


the total life, represents fatigue resistance of the joint while gives fatigue
damage over unit time or the fatigue damage rate. Apart from the inherent
randomness of the fatigue process, since the final fatigue failure event of the joint is
not defmed precisely the factor T 1. should be modelled as a random variable. The
statistics of t can be obtained from available constant amplitude experimental results
which give both number of cycles to through-thickness cracking and final failure (for
example N3 and N4 cycles reported in Dept. of Energy (1984)). The modelling of
other random variables A, K, B and B f has already been described in chapter 5.

With the above expression for base fatigue life and Eq.7.13 for fatigue damage ratio,
the actual failure times t for successive joints in the failure sequence and their
conditional safety margins can be determined as in Eq.7.7. The use of efficient
FORM for system reliability analysis allows uncertainty in the basic variables to be
described by any of the random variable models and arbitrary correlation structure.
Major limitations of the earlier methods given by Stahl and Geyer (1984) and Kumar
and Karsan (1990) can be overcome in this way. The proposed method does not
require an explicit expression for fatigue life as is necessary for the previous methods.
In addition, the use of virtual distortion method significantly reduces the effort
involved in recomputing fatigue stress distributions following the failure of joints.

Some caution should be exercised in the use of the T factor. In reality, during the
final phase of crack growth beyond through-thickness cracking, the stiffness of the
joint reduces continuously. Because of this the joint will start shedding its load to
other joints before its final failure and it is very difficult to model this continuous load
re-distribution. If this load shedding before the final failure is not modelled, the
fatigue lives of remaining joints will be over estimated. The computed probability of
7-16
the failure sequence should be considered as a lower-bound to the true path
probability. If the Tif factor is not used and fatigue lives are computed corresponding
to through-thickness cracking, the computed probability will be an upper-bound.

The proposed approach for fatigue system reliability analysis has been used to analyze
an equal strength and equal load sharing parallel member system, originally studied
by Stahl and Geyer (1984). All variables were considered as lognormal and all data
used was same as in the earlier work. Identical results to those in Stahl and Geyer
(1984) were obtained. However, it should be noted that the method of Stahl and
Geyer (1984) is limited to equal strength and equal load sharing systems while the
proposed method is more general. Application of this method to the reliability
analysis of jacket structures is demonstrated in Example 1 at the end of this chapter.

7.4.3 AnaJysis Using a Fracture Mechanics Approach:

An expression for the fatigue life of a tubular joint can be similarly derived based on
the probabilistic fatigue crack propagation model presented in chapter 5. Referring
to Eq.5.8, the base fatigue life for a joint will be

ac

- Tl1dx
(7.15)
S ,., Q m vm ,.m [ltx/q]m/2
a0 Cu1.11 a 11a

- &B8'71B31E[Sm]

where the critical crack depth a for fatigue is taken as the plate thickness. As before,
represents the fatigue resistance of a joint and the fatigue damage rate. The
terms a' Ga, and Ua are all deterministic functions of crack size and have been
defined in chapter 4. The modelling of basic random variables C, a0, Kth, Bgff

and B 1 has already been discussed in chapter 5. The factor , modelled as a random
variable, represents the uncertainty in obtaining the total fatigue life of a joint from
7-17
the time to through-thickness crack formation. The conditional safety margins for
successive joint failures can be determined as in Eq.7.7 from the expressions of base
fatigue lives and fatigue damage ratio.

The fracture mechanics formulation does not provide an explicit expression for fatigue
life in terms of all basic variables and consequently cannot be studied using the
previously available methods of Stahl and Geyer (1984) and Kumar and Karsan
(1990). The use of these methods in reliability updating based on in-service
inspection results is severely restricted. The application of the proposed method is
demonstrated on a jacket structure in Example 1 at the end of this chapter.

7.5 ANALYSIS OF A SEQUENCE OF FATIGUE AND FRACTURE FAILURES

An offshore structure undergoing fatigue deterioration in time may fail by a sequence


of fatigue failures, a scenario which can be studied by the methods given in the
previous section. However, one or more tubular joints, which have fatigue cracks of
significant size, can fail by fracture if they are subject to severe load peaks, for
example during seasonal storms. A fracture failure will cause a sudden loss in
structural strength and, unlike fatigue, does not provide time for inspection and repair.
A combination of fatigue and fracture failures can be catastrophic for the platform and
should be assessed using system reliability concepts.

Based on the CEOB R6 method, see chapter 4, a tubular joint will fail by fracture if
the fracture parameter K due to applied loading exceeds the fracture resistance K.
For a given magnitude of the applied stress S a critical crack depth a, at which
fracture will occur can be determined using the expressions for K 1 and K11. These
parameters are not explicit functions of crack depth and an iterative procedure will be
normally required. If the critical crack size for the joint for a given applied load is
known the time required for an initial weld defect of size a0 to grow by fatigue to this
critical size can be determined using a fracture mechanics crack propagation model.
The computed time will therefore represent the life of the joint which fails by a
7-18
combination of fatigue and fracture effects, or simply "time to fatigue-fracture failure".
The stress intensity factor also depends on the crack shape and this complicates the
calculations. An incremental technique can be used to determine the time to failure
as given in the flow chart below.

START

Initialize:
a-a0
c-co

Crack Growth:
a-a+tia Fatigue Model
c-c+Ac

Calculate:
Fracture Model
K r ' K1

NO
cçj( r>IaK

YES (a-a0)

Calculate:
(Eq.7.15)
''R.'&.TI

STOP

Fig.7.2: Flow-chart for the calculation of time to fatigue-fracture failure


7-19
The failure of a joint increases both the fatigue stresses and the stress S due to the
peak load event in the surviving joints. Therefore, when computing the time to
fatigue-fracture failure of a joint j, following failure of (j-1) previous joints, the
updated stress due to the peak load is used in calculating T3. The additional
effect on fatigue stresses is accounted in the usual way through the factor

The time to fatigue-fracture failure T can then be used to compute conditional safety
margins for successive joints as in Eq.7.7. Note that the calculation of '1 is
independent of the type of previous joint failures. For example, one or more of the
previous (j-1) joint failures may be by pure fatigue or by fatigue-fracture. Thus, in
the same failure sequence, a combination of fatigue and fatigue-fracture failure types
could be considered. Moreover, the use of factor even in the case of
fatigue-fracture failure implies that a semi-circular fatigue crack will grow rapidly by
fracture to form a rectangular through crack which then continues to propagate by
fatigue around the intersection. The eventual failure of the joint is still by fatigue
while the effect of fracture is merely to reduce the total fatigue life of the joint.

7.6 ANALYSIS OF COMBINED EXTREME LOAD AND FATIGUE FAILURES

While the system reliability analysis method presented in the previous chapter is good
for treating ductile failure of tubular members and joints, the time-dependent reliability
formulation developed so far in this chapter caters for only fatigue and fatigue-fracture
failures. In an offshore structure a more likely scenario will be an initial failure of
one or two joints by fatigue while the significantly weakened structure may fail under
a single large wave. Ideally a method which combines both these types of failures in
a single analysis is desired and this is attempted in the following.

Consider a failure sequence k: 1,2,...,Q,Q-e-1,D+2,...,q in which the first failures are


due to fatigue or fatigue-fracture and the remaining (q-I) are quasi-static failures
occurring under a single wave. For the structure to fail all elements in the sequence
should fail and the probability of this sequence developing can be written as
7-20

((M1 3) fl... fl (M^0)) fl ((M1 )Ø) fl... n (7.16)


Pf,k - ( )

Safety margin expressions for time-dependent fatigue/fracture failure of first Q


elements can be written as

(7.17)
- +'r' ) -

while safety margin expressions for quasi-static failure of remaining (q-Q) elements
will be of the form

p-l+1,l+2,...,q

(7.18)
Let t, be the time to failure of the last fatigue/fracture failure element. Then the
remaining exposure period for the structure within which an extreme wave can occur
(leading to the collapse of the damaged structure) is t=t-t,. Consequently, the
updated stress in the above should be calculated based on an extreme wave of
reference period r and not the design life-time extreme wave as before. If the
distribution function of annual extreme wave height H 1 is Gumbel it can be shown
that the maximum wave height I-1. in a reference period of t years will also be
Gumbel distributed. The mean and standard deviation of H can be obtained as, see
for example Thoft-Christensen and Baker (1982)

- p11+La,,ln(t)
(7.19)

a -a
H1 H1

If the other environmentaJ variables such as wave period, current speed and wind
speed are modelled using conditional distributions, the conditional mean and standard
deviations for a given wave height can be easily determined. However, if these
variables are modelled using marginal distributions the parameters can be determined,
as an approximation, corresponding to their annual extremes.
7-21
In the failure sequence as formulated above, it is interesting to note that all
fatigue/fracture failures have to precede all quasi-static failures and it is not possible
to reverse the order or to have fatigue and quasi-static failure elements alternating
each other. This is because all quasi-static failure elements are assumed to be ductile
in nature or their resistances do not change with time.

With this simple modification failure sequences involving a combination of fatigue


and extreme wave failures can be analyzed. Note that if the sequence of
fatigue/fracture failures themselves lead to the collapse of the structure, the resulting
sequence will be a pure time-dependent failure sequence. On the other hand, if there
are no fatigue/fracture failures the reference wave height corresponds to a life-time
extreme wave and a pure time-independent failure sequence will be obtained. With
this approach, within a single failure-tree analysis it is possible to enumerate failure
sequences involving: (i) only fatigue/fracture failure elements, (ii) only quasi-static
failures under a single extreme wave or (iii) a combination of the two. All the failure
sequences are now defined over the life-time of the structure and thereby the
computed reliability will be the life-time reliability of the structure.

The failure-tree enumeration method described in chapter 6 can still be used with
some modifications. In addition to the selective enumeration strategy previously
defined, the selection of additional elements to be combined with the existing partial
failure path should also consider the limit-state type of the failure elements. A useful
strategy here is to assign only fatigue/fracture limit-states to fatigue sensitive joints
while all non-fatigue sensitive joints should be assigned plastic collapse limit-state for
which the effect of fatigue deterioration could be neglected. Similarly all member
failure limit-states should be ductile failure modes. Then, in branching from a node
with a ductile failure mode, new elements should only be of ductile type and all
fatigue/fracture elements should be skipped. While branching from a node with
fatigue/fracture failure mode, all element types could be considered. This approach
has the feature that all failure sequences starting with ductile elements will lead to
pure quasi-static life-time extreme load sequences while those starting with
7-22
fatigue/fracture elements will result in either pure fatigue/fracture time-dependent
sequences or mixed failure modes.

Example 1: System Reliability of a Jacket Structure Under Fatigue Conditions.

The BRITE structure studied in chapter 6 is analyzed here for the fatigue condition.
The spectral response analysis and deterministic fatigue analysis are same as reported
in example 2 of chapter 6. The attention in the following is limited to system
reliability analysis under fatigue condition.

In the present analysis only failure sequences involving pure fatigue failures are
considered. During failure-tree enumeration the S-N approach is used as it is
considerably less expensive to use compared to the fracture mechanics approach. The
failure probabilities of the dominant sequences are then updated using a fracture
mechanics approach.

Probabilistic Modelling:
The probabilistic modelling for the S-N analysis considered four random variables: the
S-N curve parameter (K), thickness correction exponent (), Miner's damage sum at
failure (), and model uncertainty in stress concentration factors (B). The statistics
used for these variables is the same as in example 2 of chapter 5. These variables are
assumed to be statistically independent of each other and independent between joints
in the structure. For reasons of simplicity the uncertainties in fatigue loading are
expressed through a single variable in the form of the model uncertainty parameter on
global response transfer function (B) which is assumed to be fully correlated all over
the structure. In the fracture mechanics analysis, the basic variables considered
include: Paris law coefficient (C), initial defect depth (a0), model uncertainty in stress
concentration factors (Bk), model uncertainty in stress intensity factors (Ba) and
model uncertainty in degree-of-bending factors (B,), all of which are assumed to be
statistically independent of each other and independent between joints. The statistics
used for these variables is the same as in example 3 of chapter 5.
7-23
Failure-tree Enumeration:

Based on initial deterministic fatigue analysis only 10 tubular joints were found to be
most critical for the structure and only these are included in the failure-tree
enumeration. The reliability indices for the individual joints are shown in Fig.7.4.
The failure-tree notation used here is the same as in Fig.6.6 except that the second
number in the box indicates the node number. The limit-state type numbers are not
indicated as only one limit-state type is used for all components. An union over all
the joint failure events gives a system reliability index at Level 1 as 1.24.

Next, starting with joint 3346/303 failure sequences are formed at Level 2 by
combining in turn with the remaining joints. After fatigue failure of a joint the
residual strength of the member is set to zero. The change in stress-state and fatigue
damage rates following previous joint failures are determined approximately in terms
of the fatigue damage ratio y. A 100-year return wave of height 22 m is used as a
reference wave for computing y as explained in section 7.4.1. The analysis is
similarly continued to higher levels and the results are shown in Fig.7.4.

It can be seen that the reliabilities increase sharply from one level to the higher level.
For the sequences considered the structure does not become a mechanism even after
5 joint failures by fatigue. However, it is possible that the structure will not be able
to withstand a large wave after such an extent of damage, although it is not explicitly
checked for in this analysis. For this reason, system failure is approximately defined
as the failure of any 5 joints in sequence. All the dominant complete failure
sequences starting with the joint 3346/303 are shown in Fig.7.4.

A similar set of failure sequences starting with 2336/306 are shown in Fig.7.4. These
sequences are also made up of essentially the same joints. Failure sequences with
joints other than the 10 critical joints result in very low failure probabilities.
The analysis is stopped at this stage although many more dominant sequences are
possible for the structure. Following a similar approach the failure probabilities of
some of the identified sequences are re-computed and the results are shown in braces.
7-24

Discussion of Results:
The results from the analysis can be summarized as in Table 7.1.

Table 7.1: Summary of results for the fatigue condition

Reliability Index for: Value

Joint most likely to fail first ndf] 1.86



First failure of any joint 1L1 1.24
Most likely complete failure path 4.28
System failure through any path 'u 3.74

The results of reliability analysis at Level 1 show very low reliabilities for some of
the joints in comparison with the reliabilities of tubular members under extreme
loading. However, the reliabilities increase sharply at higher levels which is in
contrast to the results obtained under the extreme wave condition. This can be
explained as follows.

Under extreme wave condition in which only tubular member failures are considered,
the reliability of the structure is dominated by environmental load variables which
induce high correlations between failure elements within a sequence and correlation
between failure sequences. However, under fatigue condition, the reliabilities are
dominated by material and model uncertainty variables of individual joints and show
very low correlations between failure elements. In the present analysis, correlations
between individual joint failure events were found to be about 0.45 while the
correlations between complete failure sequences were of the order of 0.60.

Because of the low correlations between failure elements the system effects are also
high. The complexity factor, given by /, is 0.66 indicating considerable system
effects. The difference between the first member failure to the system failure is also
high in comparison with the extreme wave analysis. This implies that considerable
benefit can be obtained by usmg a system reliability approach for the fatigue condition.

7-25

1.211 2.101
11241204 2134/304
(4.200)
0.107 2 520 5.044
r- .
17241204 2330/300 5040/300
'-
I.S12
3340/303 3744/307
5.142
_3_•,sO
.560 I 4221

-S
I44/3)I_1114/204
2.100 2.724
4.251
3040/300 2134/304 2134/304
-
2.202 3.125 (4.350)
2134/304 1124/104
2.223 1 020
f-
2734/304 3040/300
L - (4.732)
1.132 2.723 4 050
#.- - 9- -..
2330/300 3144/301 1144/301

j 331 2.012 3455 4.407 / 4.415



2030/300 3744/30? I724/20 24I2o%___J 3340/SOl
'-- -
1.002 2 S40 4.543
,.- 9- -,
3144/301 3040/300 0040/300
- '- -
1.030 I ISO (4 007)
r- -
3744/30? 3144/301

Fig.7.4: BRITE smicture: Dominant fatigue failure sequences


7-26
7-27
7.7 CONCLUDING REMARKS

Previously available methods for reliability analysis of offshore structures mainly


address the issue of the structure surviving an extreme wave condition while the effect
of fatigue is completely neglected. In chapter 6 an improved system reliability method
has been presented which makes it possible to account for fatigue in the form of
deterioration in strength of tubular joints. However, this approach is not completely
satisfactory as the limit-state of fatigue (through-thickness crack) is not included and
the method at best gives annual probability of failure.

Fatigue is essentially a time-dependent phenomenon and in a structure subjected to


fatigue elemental failures will occur at different points in time. At the present time
only a few studies have been reported in the literature which deal with sequence of
fatigue failures in time. However, these methods are limited to the use of S-N
approach for fatigue damage, lognormal uncertainty models for basic variables and
simple correlation structure. Numerical integration is required for probability
calculations which severely limits the applicability of the methods to large practical
problems. Since fracture mechanics models cannot be used these methods are of little
utility in incorporating inspection results for reliability updating.

In this chapter a more realistic treatment of fatigue effects is studied. A general


formulation for time-dependent system reliability analysis has been developed which
uses FORM/SORM methods for efficient probability integration. Arbitrary random
variable models and correlation structure can be used for describing uncertainty in
basic variables. The use of this formulation for analysis of sequence of pure fatigue
failures using both S-N and fracture mechanics damage accumulation models has been
presented. An approximate but efficient method for updating fatigue stress spectra
following joint failures is discussed. The methodology has been extended to study
combinations of fatigue and fracture in which joints with significant fatigue cracks can
fail by unstable fracture under multiple load peaks (for example during seasonal
storms). An essential feature of the proposed time-dependent reliability formulation
7-28

is that all elemental failures have to be brittle in nature so that elements once failed
remain in a failed state during the remaining life of the structure.

Finally an approach for combining both time-dependent and extreme load failures has
been proposed. This method allows the development of a single failure-tree for the
structure which includes: (i) pure extreme load failure sequences in which all elements
fail under a single large (life-time extreme) wave, (ii) pure fatigue/fracture failure
sequences in which brittle elemental failures occur at various points in time during the
service of the structure and (iii) mixed failure sequences in which initial failures occur
by fatigue and/or fracture at random points in time and the significantly weakened
structure then collapses under a single large wave. All failure sequences are defined
over the service life of the structure and hence life-time reliability of the structure is
obtained. The proposed method, however, does not account for the dynamics of the
structure and is limited for application to platforms in shallow to medium water depths
for which dynamic effects can be ignored.

The method developed has been used for reliability analysis of a jacket structure under
fatigue conditions. The results show very low reliabilities with respect to the failure
of single joints or sequence of two joint failures but reliability against system failure
by fatigue alone is very high. Because of low correlations between failure elements
the system effects are high and considerable benefits can be derived by using a system
reliability approach for the fatigue condition. An example involving combination of
fatigue and extreme wave conditions could not be presented as the proposed
formulation is not yet implemented in the RASOS software.
Chapter 8
CONCLUSIONS AND RECOMMENDATIONS

8.1 GENERAL

Although the uncertainty in the prediction of environmental loads and component


strengths is well recognised, the conventional design of jacket-type offshore structures
has largely been deterministic in nature. Formal application of structural reliability
methods to components of offshore structures has been attempted since the early '70s
and this has led to the introduction of LRFD design methods for jacket-type structures
in recent years. However, structural reliability methods have far greater potential if
they can be applied directly as a part of the design and assessment process.

The past decade has seen increasing research activity in the area of structural
reliability, particularly in systems analysis, and a number of methods have been
proposed for the reliability analysis of structures. Although considerable progress has
been made in this direction, the presently available methods are inherently limited in
application to only simple structures with idealized material properties and loads, and
are not yet suitable for use with large complex structures such as offshore jackets.

The overall objective of this thesis has been to improve the current status and provide
the basis for a rigorous and yet practicable reliability assessment of jacket-type
offshore structures. In particular, emphasis has been placed on the treatment of
fatigue effects in a system reliability assessment, an aspect which the existing methods
have largely ignored. The scope of the thesis has been formulated within the overall
framework of the CEC BRITE Project No. P1270 and has considerably benefitted
from the associated developments in that project. The theoretical developments and
results of example studies on a full-scale jacket structure have already been discussed
in detail under various chapters. The main contributions and important findings of the
thesis are summarized in the following.
8-2
8.2 IMPORTANT CONTRIBUTIONS

A realistic method for system reliability analysis of jacket-type offshore structures has
to take into account all the important factors which influence the performance of these
structures. Judging by the diversity of factors which affect the behaviour of jacket
structures and the intricate nature of system probability calculations, the complexity
of the subject matter of the thesis can be easily appreciated. The thesis has strived
to achieve an appropriate balance in the presentation of mechanical aspects of jacket
behaviour, uncertainty analysis, system reliability formulation and computational
aspects of solution methods, so that all necessary background is available for any
further studies on the subject.

A number of aspects of jacket behaviour, such as load and response analysis, strength
models for plastic collapse, fatigue and fracture of tubular joints, non-linear
progressive collapse analysis, are all subjects of intense ongoing research. The state
of the art in these areas has advanced by unequal levels, often ignoring interaction of
other aspects of the problem. As a result, a fully consistent mechanical model, which
incorporates recent developments, for analysis of jacket structures is not easily
available. The thesis has attempted to provide a broad review of all recent research
in these areas. Based on this review and additional new development where
necessary, a fully consistent, comprehensive and yet reasonably simple mechanical
model, which is a prerequisite for reliability assessment, has been developed for jacket
analysis. The proposed model predicts mean-value responses and component
capacities on a limit-state approach and thus differs from the conventional procedure
used in deterministic design. Wherever possible the method developed has been
compared with available experimental results or other more advanced methods. This
forms the subject of discussion in chapters 3, 4 and part of chapter 6.

A critical review of computational methods for componential and system reliability


analysis has been presented in chapter 2. The review is lmiited to FORM and SORM
techniques while simulation methods are not discussed. A detailed review of presently
available methods for time-independent system reliability analysis of structures is
8-3
given in chapter 6 and available time-dependent reliability methods are reviewed in -
chapter 7. Inherent limitations of these methods for application to offshore structures
have been identified. A new unified approach for system reliability analysis of jacket
structures under both extreme load and fatigue conditions has been proposed which
uniquely overcomes the limitations of the present methods. The proposed method has
been combined with associated developments within the CEC BRITh project so that
a fully integrated software package called RASOS has now become available for
practical application of system reliability methods in the design and assessment of
offshore jacket structures. The methodology and the software has been used for the
reliability analysis of a 6-legged shallow water jacket structure based on which a
number of important conclusions have been drawn.

In addition to developing a comprehensive and practicable procedure for reliability


assessment of jacket-type offshore structures, the thesis has made a number of specific
new contributions as outlined in the following:

0 A short-term and a long-term probability density function model for hot-spot


stress ranges have been developed to compute cumulative fatigue loading more
accurately, even for wide-banded stress spectra (see section 3.5). The
short-term model is based on the distribution of peaks of a Gaussian process
and a closed-form expression for equivalent stress-range has been derived. A
long-term model has also been developed which combines the contributions of
a large number of sea-states into a single mixed Weibull probability density
function for stress-range. This allows efficient computation of cumulative
loading during incremental and iterative fatigue crack growth calculations
accounting for stress threshold effects.

0 A fracture mechanics fatigue crack propagation model has been developed to


estimate crack growth through the thickness and at the surface of a tubular
joint (see section 4.4.4). The model uses simplified parametric formulae for
stress distributions and stress intensity factors but explicitly accounts for the
effects of weld geometry, residual stresses, crack coalescence, load shedding,
8-4

variable amplitude loading and fatigue threshold.. This allows one to study the
sensitivity of joint reliability to these uncertainties in fatigue life predictions.

O A fracture assessment model for tubular joints has been developed based on
the CEGB R6 approach (see section 4.5 2). The plastic collapse parameter is
derived from limit-load model for cracked joints and the fracture parameter
considers through thickness distribution of primary and secondary stresses.

O A detailed discussion on sources of uncertainties in extreme environmental


loading, fatigue loading, material properties and mechanical models for
response and strength analysis has been presented (see sections 5.2-5.5). A
number of alternative methods for quantification of these uncertainties in a
reliability analysis have been proposed.

O Reliability analysis models for tubular joints have been presented for the
limit-states of plastic-collapse, fatigue and fracture (see sections 5.6-5.8). The
uncertainties in material parameters: such as yield stress, fracture toughness,
crack growth parameters, weld defect size, residual stresses; loading
parameters: such as wave height, period, current and wind, drag and inertia
coefficients; and model uncertainty parameters: such as stress concentration
factors, stress intensity factors, response transfer function, collapse loads etc.
have been considered. Probability of failure and sensitivity factors have been
calculated using FORM methods.

o A new system reliability method for the extreme wave condition of a jacket
structure has been developed. The method uses a response surface approach
for environmental load modelling and a virtual distortion method (VDM) of
non-linear structural analysis developed jointly with others under the BRFE
Project P 1270. A joint beta-point concept based on multi-constraint
optimization has been used for reliability analysis of a prescribed failure
sequence (see sections 6.5.3 and 6.5.4). This avoids the problem of physically
inadmissible design points of the crude FORM approach and the limitation of
8-5
strain-independent post-limit models of the member replacement approach used
by earlier methods. An approach has also been proposed for incorporating
fatigue effects in a extreme wave reliability analysis by treating it as a
deterioration in tubular joint strength under plastic collapse and fracture modes
(see section 6.7).

O A new method of failure-tree enumeration has been developed which uses


selective enumeration and preferential branching to identify failure sequences
in the order of their importance and computes bounds on the system reliability
to within user specified limits (section 6.6.3). The method is to some extent
approximate but is highly efficient compared to the previously available
rigorous methods. Combined with the response surface approach for wave
loading and the virtual distortion method of non-linear analysis the proposed
method makes a practical reliability analysis of jacket structures possible.

O A general formulation for time-dependent system reliability has been


developed for analysis of sequence of pure fatigue failures, using both S-N and
fracture mechanics damage accumulation models (sections 7.3 and 7.4). An
approximate but efficient method for updating fatigue stress spectra following
joint failures is discussed. The methodology has been extended to study
combinations of fatigue and fracture (section 7.5) in which joints with
significant fatigue cracks can fail by unstable fracture under multiple load
peaks (for example during seasonal storms).

O Finally an approach for combining both time-dependent and extreme load


failures has been proposed (section 7.6). This method allows the development
of a single failure-tree for the structure which includes: (i) pure extreme load
failure sequences in which all elements fail under a single large (life-time
extreme) wave, (ii) pure fatigue/fracture failure sequences in which brittle
elemental failures occur at various points in time during the service of the
structure and (iii) mixed failure sequences in which initial failures occur by
fatigue and/or fracture at random points in time and the significantly weakened
8-6
structure then collapses under a single large wave. All failure sequences are
defined over the service life of the structure and hence life-time reliability of
the structure is obtained.

The proposed method, however, uses only random variable uncertainty models and
does not account for the dynamic effects of the structure. The method is therefore
applicable to jacket structures in shallow to moderate water depths.

8.3 SUMMARY OF FINDINGS

The main objective of the thesis has been to develop new and efficient methods for
reliability assessment of jacket platforms, in which modest success has been achieved.
The method developed has been implemented into an operational computer software.
Within the available time it has not been possible to apply the developed method to
full-scale analysis of jacket structures but this should be possible. The computational
and human time required for such analysis is expected to be quite high. It is proposed
that the potential of the method developed here for practical application in the design
and assessment of offshore structures should be evaluated by carrying out a number
of full-scale case studies on jackets. Efforts are already under way to establish a Joint
Industry Project on these lines. However, a limited analysis has been carried out on
a 6-legged shallow-water jacket structure based on which a number of conclusions can
be drawn as summarized below.

Reliability analysis has been carried out for various joints in the structure for the
limit-states of fatigue, plastic collapse and fracture (sections 5.6-5.8) based on which
following conclusions have been drawn.

- The results for fatigue limit-state using an S-N approach indicate that the
reliability of the joint is typically highly sensitive to model uncertainties in the
transfer functions for hydrodynamic loading, response analysis for nominal
member stresses and stress concentration factors. The uncertainty in S-N data
8-7
and the assumed value of Miner's damage sum at failure are also important.
The reliability is seen to be less sensitive to uncertainties in the stress
endurance limit and thickness correction factor.

In a fracture mechanics based reliability assessment the model uncertainties in


calculating the degree of bending factors and stress intensity factors also
contribute significantly. The reliability of the joint is also sensitive to
uncertainties in initial defect size and crack growth material parameters. With
the median fatigue life being the same and for the same service exposure, the
fracture mechanics approach predicts a lower reliability for a joint compared
with the S-N approach. This is because of the additional uncertainties
involved in computing the degree of bending and stress intensity factors using
parametric formulae. However, when these uncertainties are reduced by
careful modelling the reliability of the joint can be significantly improved. For
this reason it is felt that at least the stress distribution of the intact joint should
be derived using finite element methods for use in fracture mechanics fatigue
assessments.

- The reliability of the joint under plastic collapse mode is governed by the
uncertainties in extreme wave loading and the model uncertainty in collapse
load calculations. Yield strength of the material has relatively little influence
because of its low variability. Inclusion of random fatigue crack growth over
an exposure period of up to 20 years did not show any influence on joint
reliability under plastic collapse for most joints.

Results for the fracture limit-state depend on many factors and in the example
considered varied for different joints in the structure. If the direction of
approach of the extreme wave is such that high tensile stresses are introduced
in joints which are critical in fatigue, the reliabilities of these joints were seen
to be sensitive to both fatigue and fracture variables and the reliabilities reduce
with exposure time. In such cases reliabilities also depend on the variation of
fracture toughness within the HAZ and extent of HAZ below the weld.
8-8

For other joints the reliability was seen to be strongly influenced by mean
value of residual stresses. For as-welded joints, with mean tensile residual
stresses in the order of 300 MPa, the reliability is highly sensitive to material
fracture toughness, residual stresses and model uncertainty in stress intensity
factors. Model uncertainties in stress concentration factors and degree of
bending did not show significant influence as the joint failure was dominated
by secondary stresses. For stress-relieved joints, on the other hand, the joint
failure was associated with high plasticity and consequently reliability was
more sensitive to uncertainties in extreme loading variables and relatively less
sensitive to fracture toughness and model uncertainty in stress intensity factors.
The reliability index in such cases was close to that under a plastic collapse
failure mode.

For all the joints analyzed, those which were critical in fatigue showed the
lowest reliabilities. The effect of fracture on joint reliability was highest only
when combined with significant fatigue. When fatigue is ignored, reliabilities
under fracture mode were found to be very high even for joints in which mean
value of initial defect depth was up to 1mm. For stress-relieved joints the
dominant mode is plastic collapse.

From the above results it appears that for tubular joints usually one of the
three modes is probabilistically dominant and different joints in the structure
are critical under different modes. To reduce computational time and to avoid
convergence problems during system analysis it is suggested that a preliminary
reliability analysis be carried out for various joints in the structure and based
on which the joints should be categorised as fatigue-critical, collapse-critical
and fatigue-fracture-critical. Then during a system analysis corresponding
limit-states and basic variables should be assigned to these joint groups.
8-9
Based on the system reliability analysis of the BRJTE structure the following
conclusions can be drawn. These conclusions are, however, limited to the particular
structure studied and cannot be easily generalized to other structures.

Under extreme wave condition the reliability of the structure is dominated by


the uncertainty in extreme environmental loading variables. Because of high
loading uncertainties the failure events of individual elements and those of
complete failure sequences are highly correlated.

For the example structure considered, the failure sequence identified from a
deterministic collapse analysis is also seen to be the probabilistically
most-likely failure sequence.

- The example structure shows high reserve against first member failure but the
redundancy beyond first failure is relatively low.

- Reliability analysis of the structure considering fatigue effects in the form of


deterioration in tubular joint strength did not show much influence of fatigue
on the system reliability.

System reliability analysis of the structure considering sequence of fatigue


failures (through-thickness cracking) shows that reliabilities against first joint
failure or a sequence of two joint failures is relatively low. However,
reliabilities increase sharply at higher levels involving more than two joint
failures which is in contrast to the extreme wave analysis.

- The system reliability under fatigue condition is dominated by resistance


uncertainties with low correlation between failure events. Because of this the
system effects are higher under fatigue condition compared to the extreme
wave condition. Because of this, considerable benefits can be derived by using
system reliability approach for fatigue effects.
8-10
8.4 POTENTIAL APPLICATIONS

Considering the large number of uncertainties which influence the reliability of jacket
structures it seems prudent to use system reliability methods for routine design and
assessment of offshore jacket structures. However, as these methods are complex and
are expensive in computer and human resources and owing to the lack of software and
trained manpower, it is unlikely that they will be used for routine designs in the
immediate future. Nevertheless, a number of special areas can be identified for
immediate application where the use of system reliability methods developed in this
thesis can give considerable benefits in terms of safety and economy. These include:

a Evaluation of new platform concepts for which previous experience is not


available. A number of alternatives can be compared in terms of their system
reliability, redundancy and damage tolerance.

* Safety assessment of existing and older platforms which do not conform to


present code requirements or when extension of platform life is being
reviewed. The available information regarding past performance and additional
detailed underwater surveys could be accommodated in a reliability analysis.

a Planning of in-service inspection and maintenance programs for fatigue. The


initial choice of joints for inspection could be based on the reliabilities of
individual joints and the effect the failure of these joints would have on system
reliability. Inspection intervals can be optimized from the information of
reliability variation with service. These, in addition, will need Bayesian
reliability updating techniques using the results of inspections.

To assess the implications of any inadvertent damage which the structure may
suffer during its service life either due to boat impacts or severe storms.
Important decisions about the continued use of the platform or the need for
expensive repairs should be based on detailed reliability evaluation.
8-11
8.5 SUGGESTIONS FOR FURThER RESEARCH

The proposed reliability analysis procedure provides a modular framework covering


all important aspects of the problem so that future improvements to particular elements
can be easily incorporated. In the following, a few important areas are identified
which need further research to enhance the applicability of the system reliability
methods to offshore structures.

A Response Surface Approach for Fatigue Loading:


Spectral response analysis methods used for computing fatigue loading are very
expensive for inclusion in iterative reliability calculation. Development of an efficient
response surface model would be highly useful in the explicit modelling of all fatigue
loading variables which is important for the system reliability analysis of jackets.

Improved Failure-Tree Enumeration Techniques:


Considerable computational time in a system reliability analysis is devoted to
identification of dominant failure paths. Efficiency of the proposed approach can be
significantly enhanced by using limited importance sampling around the joint
beta-point to select important sequences for enumeration from the selected node.

System Reliability Updating Based on Inspection Results:


At present, Bayesian methods are available for updating reliability of individual joints.
Methods for updating system reliability based on results of inspections are needed to
apply the proposed method for in-service inspection planning and decisions on repair.

Extension to Dynamically Sensitive Offshore Structures:


Extension of the proposed method to account for dynamic effects is tequired for
application to deep water jackets and compliant structures. A long-term reliability
method should be preferred in which a random process model for the time-varying
loading to compute mean out-crossing rates could be used in conjunction with a
time-independent reliability analysis in the space of uncertain resistances. New
methods are also required for progressive collapse analysis under dynamic loading.
REFERENCES

Aaghaakouchak, A., Glinka, G. and Dharmavasan, S. (1989): "A load shedding model
for fracture mechanics analysis of fatigue cracks in tubular joints", Proc. Offshore
Mechanics and Arctic Engineering Confemce, The Hague, Netherlands.

Abbassian, F., Lau, T.B. and Zintilis, G.M. (1989): "3-D inelastic behaviour of tubular
members supported on non-linear rotational springs", Offshore Mechanics and Arctic
Engineering Conference, The Hague, Netherlands.

Abramowitz, M. and Stegun, LA. (1965): "Handbook of mathematical functions",


Dover, New York.

Ainsworth, R.A. (1984): "The assessment of defects in structures of strain hardening


material", Engineering Fracture Mechanics, Vol 19, No 4.

Ainsworth, R.A. et al (1986): "Revision 3 of R6, its background and validity", Proc.
6th European Conference on Fracture, ECF6, Vol.2, Amsterdam, The Netherlands.

Akiniwa et al (1988): "Propagation and closure of short cracks at notches under


compressive mean stresses", Fatigue and Fracture of Engineering Materials and
Structures, Vol.11, No.5.

Albrecht, P. and Yamada, K. (1977): "Rapid calculation of stress intensity factors",


J. of Structural Divn., ASCE, Vol.103, No.ST2.

American Petroleum Institute (1987): "Recom,nended Practice for Planning, Designing


and Constructing FLIed Offshore Platforms", API RP 2A, 17th Edn, Washington, DC.

American Petroleum Institute (1989): "Draft recommended practice for planning,


designing and constructing fixed offshore plafor,ns - load and resistance factor
design", API RP 2A-LRFD, First Edn.

American Society for Testing of Materials, (1979): "Standard test method for plane
strain fracture toughness of metallic materials", ASTM E399-78a.

Anderson, T. L., Leggat, R.H. and Garwood, Si. (1988): "The use of CTOD methods
in fitness for purpose analysis", European Conference on Fracture.

Badshauge, 0. and Bach-Gansmo, 0. (1985): "System reliability analysis of a jacket


structure", Proc. 4th mt. Conf. on Structural Safety and Reliability, ICOSSAR.

Baker, M.J. (1972): "Variability in the strength of structural steels - A study in


structural safety, Part I; Material variability", CIRIA Technical Note 44.

Baker, Mi. and Wyatt, T. (1979): "Methods of reliability analysis for jacket
platforms", Proc. Behaviour of Offshore Structures Conference, BOSS-79, London.
R-2
Baker, M.J. and Ramachandran, K. (1981): "Reliability analysis as a tool in the design
of fixed offshore platforms", Proc. 2nd mt. Symposium on the Integrity of Offshore
Structures, Glasgow, U.K.

Baker, M.J. (1985): "The reliability concept as an aid to decision making in offshore
structures", Proc. Behaviour of Offshore Structures Conference, BOSS-85, Amsterdam,
The Netherlands.

Baker, M.J., Kountoris, I.S. and Ohmart, R.D. (1988): "Weld defects in an offshore
structure - A detailed study", Proc. Behaviour of Offshore Structures Conference,
BOSS-88, Trondheim, Norway.

Bea, R.G., Pawsey, S.F. and Litton, R.W. (1988): "Measured and predicted wave
forces on offshore platforms", Proc. Offshore Technology Conference, OTC 5787,
Houston, Texas.

Berge, S. (1985): "On the effect of plate thickness in fatigue of welds", Engineering
Fracture Mechanics, VoL21, No.2.

Berge, S. et al (1987): "Effect of plate thickness in fatigue of welded joints in air and
in sea water", Proc. 3rd. Int. Conf. on Steel in Marine Structures, Deift.

Berranger, I., Goyet, J. and Gierlinski, J.T. (1991): "Reliability analysis of an offshore
structure using RASOS software", Proc. Conf. Brazil Offshore.

Bilby, B.A., Cottrell, A.H. and Swindon, K.H. (1963): "The spread of plastic yield
from a notch", Proc. of the Royal Society, A272.

Birades, M. and Shetty, N.K. (1991): "Computation of hot-spot stress distribution",


RASOS-H: User's Manual.

Bjerager, P. et.al. (1988): "Reliability method for marine structures under multiple
environmental load processes", Proc. Conf. Behaviour of Offshore Structures,
BOSS-88, Trondheim, Norway.

Bjerager, P. and Krenk, S. (1989): "Parametric sensitivity in first order reliability


theory", ASCE, J. of Eng. Mech., Vol. 115, No. 7.

Bjerager, P. (1990): "On computation methods for structural reliability analysis",


Structural Safety, Vol.9.

Bogdanoff, J.L. and Kozin, F. (1985): "A Probabilistic Approach to Cunudative


EJamage"john Wiley and Sons, New York, NY.

Breitung, K. (1984). "Asymptotic approximate for multinormal inregrals", ASCE, J.


of Eng. Mech., Vol. 110, No.EM3.
R-3
British Standards Institution, (1977): "Methods of test for plane strain fracture
toughness (Krn) of metallic materials", BS5447, London, U.K.

British Standards Institution, (1979): "Methods for crack opening displacement


testing", BS5762, BSI, London.

British Standards Institution, PD6493, (1980): "Guidance on some methods for the
derivation of acceptance levels for defects in fusion welded joints", BSI, London, U.K.

British Standards Institution, (1989): "Draft piblished document for guidance on some
methods for the derivation of acceptance levels for flaws in fusion welded joints,
(Revision of PD6493)", BSI, London, U.K.

Burdekin, F.M. and Stone, D.E.W. (1966): "The crack opening displacement approach
to fracture mechanics in yielding", Journal of Strain Analysis, Vol.1.

Burdekin, F.M. et al (1988): 'Design assessment support for tubulars by fracture


mechanics methods", Proc. Conf. on Fatigue in Offshore Structures, (eds) Wi). Dover
and 0. Glinka, London, U.K.

Carr, P. and Birkinshaw, (1988): "Development of a methodology to make use of the


results of sensitivity studies for analysis of uncertainties in extreme environmental
loading on offshore structures", 011 537, HMSO, London, U.K.

Castro, D.E., Marci, G. and Munz, D. (1987): "A generalised concept of a fatigue
threshold", J. of Fatigue and Fracture of Engineering Materials and Structures",
Vol.10.

Chell, 0.0. (1976): "Bilby, Cotirell and Swindon model solutions for centre and edge
cracked plates subjected to arbitrary mode I loading", Int. J. of Fracture, Vol 12, No.1.

Choudhury, G. and Dover, W.D. (1985): "Fatigue analysis of offshore platforms


subjected to sea wave loadings", Int. J. of Fatigue, Vol. 7, No. 1.

Connolly, M.P. et al (1989): "A parametric study of the ratio of bending to membrane
stress in tubular Y and T joints", Paper submitted to the Journal of Strain Analysis for
Engineering Design.

Cornell, C.A. (1969). "A probability-based structural code", J. of the American


concrete Institute, Vol. 66, No.12.

Dawes, M.G. (1980): "The COD design curve", Advances in Elasto-plastic Fracture
Mechanics, ed. Lii Larsson, Applied Science Publishers, London, U.K.

de Koning, A.U. (1981): "A simple crack closure model for prediction of fatigue crack
growth rates under variable-amplitude loading", ASTM, STP 743.
R-4
Deift (van), D.R.V., Dijkstra, O.D. and Snijder, H.H. (1986): "The calculation of
fatigue crack growth in welded tubular joints using fracture mechanics", Offshore
Techonology Conference, Paper No. OTC 5352, Houston, Texas.

Delmer, M.V., Sorensen, J.D. and Thoft-Christensen, P. (1988): "Collapse probability


for elasto-plastic beam structures", Proc. Second IFIP WP 5.7 Working Conference,
London, U.K.

Department of Energy (1984): "Background to new fatigue design guidance for steel
welded joints in offshore structures", HMSO, London, U.K.

Department of Energy (1987): "U.K. offshore steels researvh project-Phase II",


0Th 225, HMSO, London, U.K.

Department of Energy (1988): "U.K. offshore steels research project-Phase I",


0Th 262, HMSO, London, U.K.

Department of Energy (1989a): "Static strength of large scale tubular joints:


Engineering assessment", 0TH 297, HMSO, London, U.K.

Department of Energy (1989b): "Background to new static strength guidance for


tubular joints in steel offshore structures", 0TH 308, HMSO, London, U.K.

Department of Energy (1990): "Offshore installations: Guidance on design,


construction and certification", HMSO, London, U.K.

Der Kiureghian, A. and Liu, P. (1986). "Structural reliability under incomplete


probability information", J. of Eng. Mech., ASCE, Vol. 112, No. 1.

Dharmavasan, S. and Dover, W.D. (1984): "Stress distribution formula and comparison
of three stress analysis techniques for offshore structures", Proc. Offshore Mechanics
and Arctic Engineering Conference, New Orleans.

Dharmavasan, S. and Dover, WI). (1988): "Nondestructive evaluation of offshore


structures using fracture mechanics", Applied Mechanics Reviews, Vol.4 1, No.2.

Dijkstra, O.D., Snijder, H.H. and Van Rongen, H.J.M. (1990): "Assessment of the
remaining fatigue life of defective welded joints", Proc. Conf. mt. Association of
Bridge and Structural Engineers", JABSE-90.

Ditlevsen, 0. (1973). "Structural reliability and the invarience problem", Research


Report, No. 22, Solid Mechanics Division, University of Waterloo, Canada.

Ditlevsen, 0. (1979a). "Generalized second moment reliability index", J. of Structural


Mech.,Vol. 7, No. 4.

Ditlevsen, 0. (1979b). "Principle of Normal Tail Approximation", J.of. Eng. Mech.


D v. ASCE, Vol.107, No EM6
R-5
Ditlevsen, 0. (1981). "Narrow reliability bounds for structural systems", I. of Eng.
Mech., ASCE, Vol. 107. No. EM6.

Ditlevsen, 0. and Bjerager, P. (1984): "Reliability of highly redundant plastic


structures", ASCE, Journal of Engineering Mechanics, Vol.110, No.5.

Dolinski, K. (1983): "First-order second-moment approximation in reliability of


structural systems: critical review and alternative approach", Structural Safety, Vol.1.

Dover, W.D. and Dharamavasan, S. (1982): 'Fatigue fracture mechanics analysis of


T and Yjoints", Papaer OTC 4404, Offshore Technology Conference, Houston, Texas.

Dover, WI). et al (1988): 'Fatigue crack growth in X joints and multi-brace nodes",
Fatigue of Offshore Structures Conference, (Eds) W.D. Dover and 0. Glinka, London.

Dover, WI)., Kare, R.F. and Hall, M.S. (1991): "The reliability of SCF predictions
using parametric equations: A statistical analysis", Proc. Offshore Mechanics and
Arctic Engineering Conference, Vol. 111-B, Stavanger, 1991.

Dowling, N.E. (1972): "Fatigue prediction for complicated stress-strain histories", J.


of Materials, JMASA, 7.

ECSC (1987): "Steel in Marine Structures", Conference Proceedings, (eds) Dc Back


and Noordhoek, C., Elsevier.

Efthymiou, M. (1988): 'Development of SCF formulae and generalised influence


functions for use in fatigue analysis", Conf. on Recent Developments in Tubular Joint
Technology - OTJ'88, Surrey, U.K.

Elber, W. (1971): "The significance of crack closure", ASTM STP 486.

El-Haddad, M.H., Topper, T.H. and Topper, T.N. (1981); "Fatigue life predictions of
smooth and notched specimens based on fracture mechanics", Journal of Engineering
Materials and Structures, Vol.103.

Effiott, K.S. and Fessler, H. (1986): "Stresses at weld toes on non-overlapped tubular
joints", Paper C131/86, Proc. Conf. on Fatigue and Crack Growth in Offshore
Structures, IMechE-86, London, U.K.

Ewalds, HI. and Wanhill, R.J.H. (1985): "Fracture Mechanics", Edward Arnold,
London, U.K.

Forman, R.G. et al (1964): "Numerical analysis of crack propagation in cyclic loaded


structures", AS ME, Paper No. 66-WA/Met 4.

Frangopol, D.M., lizuka, M. and Yoshida, K. (1991): "Redundancy measures for


design and evaluation of structural systems", Offshore Mechanics and Arctic
Engineering Conference, Stavanger, Norway.
R-6

Gibstein, M. and Moe, E.T. (1986): "Brittle fracture risks in tubular joints", Offshore
Mechanics and Arctic Engineering Conference, Houston, Texas.

Gierlinski, J.T. and Lau, B. (1991): "Non-linear analysis using the virtual distortion
method", RASOS_C Programmers Manual.

Gill, P. Murray, W. and Wright, M. (1981). "Practical optimization",Academic press,


London.

Gollwitzer, S. and Rackwitz, R. (1983). "Equivalent components in first-order system


reliability", Reliability Eng. Vol. 5.

Gollwitzer, S. and Rackwitz, R. (1988). "SYSRELICUTALG manual", RCP, GmbH,


Munich.

Gollwitzer, S.; Guers, F. and Rackwitz, R. (1988). "SORM manual", RCP, GmbH,
Munich.

Gollwitzer, S. (1991): "Reliability analysis by failure-tree enumeration", RASOS_R


Theoretical Manual.

Goyet, J., Mariem, B. and Shetty, N.K. (1991): "System Reliability of offshore
structures using failure-tree approach: RASOS_R Theoretical Manual".

Guenard, Y. (1984): "Application of system reliability anlysis to offshore structures",


John Blume Earthquake Engineering Centre, Report No.7 1, Stanford.

Guers, F., Dolinsld, K. and Rackwitz, R. (1988): "Probability of failure of brittle


redundant structural systems in time", Structural Safety, Vol.5.

Gulati, 0. et al (1982): "An analytical study of stress concentration effects in multi-


brace joints under combined loading", Proc. Offshore Technology Conference, Paper
No. OTC 4407, Houston.

Hancock, J.W., Gall, D.S. and Huang, X. (1986): "Fatigue crack growth due to
random loading", Proc. Fatigue and Crack Growth in Offshore Structure Conf.,
IMechE-86, Institution of Mechanical Engineers, London.

Hasofer, A.M. and Lind, N.C. (1974). "Exact and invariant second moment code
format", J. of Eng. Mech. Div., ASCE, Vol.107.

Haswell, J. (1991): "A defect assessment methodology for offshore jacket structures
including complex joints and complex loading", Conf. on Integrity of Offshore
Structures, IGS-90, Glasgow, U.K.

Haver, S. and Natwig, B. (1991): "On some uncertainties in the modelling of ocean
waves and their effects on TLP response", mt. J. of Offshore and Polar Engineering,
Vol. 1, No. 2.
R-7
Haver, S. and Winterstein, S.R. (1990): "The effects of joint description of
environmental data on design loads and reliability", Proc. 9th mt. Conf. on Offshore
Mechanics and Arctic Engineering, Vol. 2, Houston.

Hayes, D.J. and Williams, J.G. (1972): "Practical method for determining Dugdale
model solutions for cracked bodies of arbitrary shape", hit. J. of Fracture Mechanics,
Vol 8.

Heiot and Benoit (1990): "Stability of surface cracks in tubular joints", BRITE
P1270, Internal Report

Hellier, A.K., Connolly, M.P. and Dover, W.D. (1990): "Stress concentration factors
for tubular Y and T joints", mt. Journal of Fatigue, Vol 12, No 1.

Heyman, J. (1971): "Plastic design of frames", VoL2, Cambridge University Press.

Hibberd, RD. and Dover, WD. (1977): "Random load fatigue crack growth in T
joints", Paper No. OTC 2853, Offshore Technology Conference, Houston, Texas.

Hohenbichler, M. and Rackwitz, R. (1981): "Non-normal dependent vectors in


structural reliability", J.of Eng. Mech. Div., ASCE, Vol.107.

Hohenbichier, M. and Rackwitz, R. (1983): "First-order concepts in system reliability",


Structural Safety, Vol.1.

Hohenbichler, M. and Rackwitz, R. (1986): "Sensitivity and importance measures in


structural reliability", Civil Engineering Systems, Vol.3.

Hohenbichier, M. et al. (1987): "New light on first-and second-order reliability


methods", Structural Safety, Vol.4.

Hoim, C.A., et al (1988): "System reliability of offshore jacket structures by elastic


and plastic analysis", Proc. Behaviour of Offshore Structures, BOSS-88.

Holnicki-Szulc, J. and Gierlinski, J.T. (1989): "Structural modifications simulated by


virtual distortions", International Journal for Numerical Methods in Engineering,
Vol.28.

Huang, X. and Hancock, J.W. (1988): "The stress intensity factors of semi-elliptical
cracks in a tubular welded T-joint under axial loading", Engineering Fracture
Mechanics, Vol.30, No.1.

Inglis, R.B., Pijfers, J.G.L and Vugts, J.H. (1985): "A unified probabilistic approach
to predicting the response of offshore structures, including the extreme response",
Behaviour of Offshore Structures, Elsevier, Amsterdam.

Johnson, N.L. and Kotz, S. (1972): "Distributions in Statistics: Continuous


Multivariate Distributions", John Wiley and Sons, New York.
R-8

Kam, J.C.P. et al (1987): "Fracture mechanics modelling and structural integrity of


welded tubular joints in fatigue", Offshore Mechanics and Arctic Engineering
Conference, Houston, Texas.

Kam, J.C.P. and Dover, W.D. (1988): 'Past fatigue assessment procedure for offshore
structures under random stress histoiy", Institution of Civil Engineers, Paper 9333,
Part 2, Vol. 85.

Kam, J.C.P. and Dover, W.D. (1989): "Advanced tool for fast assessment of fatigue
under random wave stress histories", Proc. Institution of Civil Engineers, Paper 9475,
Part 2, 87.

Karadeniz, H. (1989): "Advanced stochastic analysis program for offshore structures",


Deift University of Technology.

Karamchandani, A. (1990): "Limitations of some of the approximate structural analysis


methods that are used in structural system reliability", Structural Safety, Vol.7.

Karamchandani, A., Dalane, J.I. and Bjerager, P. (1991): "Systems reliability of


offshore structures including fatigue and extreme wave loading", Marine Structures,
Vol.4.

Karsan, D.I. and Kumar, A. (1990): "Fatigue failure paths for offshore platform
inspection", ASCE, Journal of Structural Engineering, Vol.116, No.6.

Kirkemo, F. (1988): "Application of probabilistic fracture mechanics to offshore


structures", Applied Mechanics Reviews, Vol. 41, No.2. Also in Fracture Mechanics
in Offshore Industry, ASME Book No. AMRO32.

Kiesnil, M. and Lucas, P. (1972): "Influence of strength and stress histoiy on growth
and stabilisation of fatigue cracks", Engineering Fracture Mechanics, Vol. 4.

Kountoris, I.S., and Baker, M.J. (1989): "Defect Assessment - Analysis of defects
detected by MPI in an offshore structure", CESUC Report No. 0R6, Dept. of Civil
Engineering, Imperial College, London, U.K.

Kuang, J.G., Potvin, A.B. and Leick, R.D. (1975): "Stress concentrations in tubular
joints", Proc. Offshore Technology Conference, Paper No. OTC 2205, Houston.

Kumar, A. and Karsan, D L (1990): 'Fatigue reliability of parallel systems", ASCE,


Journal of Structural Engineering, Vol.116, No.3.

Labeyrie, J., Cazzulo, R. and Dogliani, M. (1992): "Multivanate load modelling for
jacket platforms reliability", Structural Safety (to be published).

Lalani, M., Tebbett, I.E. and Choo, B.S. (1986): 'Improved fatigue life estimation of
tubular joints", Proc. Offshore Technology Conference, OTC 5306, Houston, Texas.
R-9
Lalani, M. and Shuttleworth, E.P. (1990): 'The ultimate limit-state of offshore
platforms using reserve and residual strength principles", Offshore Technology
Conference, OTC 6309, Houston, Texas.

Langley, R.S. (1989): "Stochastic models for crack growth", Engineering Fracture
Mechanics, Vol. 32.

Lin, Y.K. and Yang, J.N. (1983): "On statistical moments of fatigue crack
propagation", Engineering Fracture Mechanics, Vol. 18.

Lloyd, J.R. and Clawson, W.C. (1984): "Reserve and residual strength of pile founded
offshore platforms", The Role of Design, Inspection and Redundancy in Marine
Structural Reliability, National Academy Press, Washington D C.

Longuet-Higgins, M.S. (1983): "On the joint distribution of wave periods and
amplitudes in a random wave field", Proc. of Royal Society, London, Vol.389.

Machida, S., Hagiwara, Y. and Kajimoto, K. (1987): "Evaluation of brittle fracture


strength of tubular joints of offshore structures", Offshore Mechanics and Arctic
Engineering Conference, Houston, Texas.

Madsen, H.O. (1986): "Extreme-value statistics for non-linear stress combination",


ASCE, J. of Eng. Mech., Vol. 111, No. 9.

Madsen, H.O., Krenk, S. and Lind, N.C. (1986): "Methods of Structural Safety",
Prentice-Hall Inc., Englewood Cliffs, N.J.

Madsen, H.O. (1987): "Model updating in reliability theory", Proc. ICASP-5,


Vancover, Canada.

Madsen, H.O. (1988): "Omission sensitivity factors", Structural Safety, Vol.5.

Maihotra, A.K. and Penzien,J. (1970): "Non-deterministic analysis of offshore


structures", ASCE, Journal of Eng. Mech. Div., EM6.

Marshall, P.W. and Luyties, W.H. (1982): "Allowable stresses for fatigue design",
Proc. Conf. Behaviour of Offshore Structures, BOSS-82, Vol. 2, MIT.

Martin, T. (1978): "The fatigue strength of welded tubular T joints with a large
diameter ratio", Paper 32, Seminar on European Offshore Steels Research, Cambridge.

Martindale, S.G. and Wirsching, P.H. (1983): "Reliability based progressive fatigue
collapse", ASCE, Journal of Structural Engineering, Vol.109, No.8.

Mattheck, C., Morawietz, P. and Munz, D. (1983): "Stress intensity factor at the
surface and at the deepest point of semi-elliptical surface crack in plates under stress
gradients", mt. J. of Fracture, Vol.23.
R-10

Meichers, R.E. and Tang, L.K. (1984): "Dominant failure modes in stochastic
structural systems", Structural Safety, Vol.2.

Mime, L et al (1985): "Assessment of the intgrity of structures containing defects",


C.E.G.B. R/HR6, Draft 4.

Moses, F. (1982): "System reliability developments in structural engineering",


Structural Safety, Vol.1.

Moses, F. (1990): "New directions and research needs in system reliability research",
Structural Safety, Vol.7.

Murotsu, Y. et al (1980): "Reliability analysis of truss structures by using matrix


method", Transactions of the ASME, Journal of Mechanical Design, Vol.102, No.4.

Murotsu, Y. et al (1981): "Reliability assessment of redundant structure", Structural


Safety and Reliability (eds. T. Moan and M. Shinozuka), Elsevier.

Murotsu, Y. et al (1984): "Automatic generation of stochastically dominant modes of


structural failure in frame structures", Structural Safety, Vol.2.

Murotsu, Y. et al (1985): "Probabilistic collaspe analysis of offshore structures", Proc.


Offshore Mechanics and Arctic Engineering Conference, Houston, Texas.

Nafde, A.M., Corotis, R.B. and Cohon, J.L. (1987): "Failure mode identification for
structural frames", ASCE, J. of Structural Engineering, Vol. 113, No.7.

Nerzic, R. and Lebas, 0. (1988): "Uncertainties in wave loading from full-scale


measurements", Proc. Behaviour of Offshore Structures Conference, BOSS-88,
Trondheim, Norway.

Newland, D.E. (1975): "An introduction to Random vibration and spectral analysis",
Longman, London.

Newman, J.C. and Raju, I.S. (1981): "An empirical stress intensity factor equation for
the sw-face crack", Engineering Fracture Mechanics, Vol.15.

Nicholson, R.W. (1987): "An elastic-plastic line-spring model for the analysis of
cracks in tubular structures", Offshore Mecchanics and Arctic Engineering Conference,
Houston, Texas.

Niu, X. and Glinka, 0. (1987): "The weld profile effect on stress intensity factors in
weidments", mt. I. of Fracture, Vol.35.

Niu, X. and Glinka, 0. (1989): "Stress intensity factors for semi-elliptical surface
cracks in welded joints", Irn J. of Fracture, Vol.37.
R-11
Nolte, K.G. and Hansford, J.E. (1976): "Closed form expressions for determining the
fatigue damage of structures due to ocean waves", Proc. Offshore Technology Conf.,
Paper OTC 2606, Houston.

Nordal, H., Cornell, C.A. and Karamchandani, A. (1987): "A structural system
reliability case study of an 8-legged steel offshore production platform", Marine
Structural Reliability Symposium, Arlington, Virginia.

Norwegian Petroleum Directorate, (1989): "Regulation for the Structural Design of


Load-bearing Structures Intended for Exploitation of Petroleum Reserves", NPD,
Stavanger, Norway.

Olufsen, A. and Bea, R.G. (1990): "Loading uncertainties in extreme waves", Marine
Structures, Vol 3.

Palmberg, B., Blon, A.D. and Eggwertz, S. (1986): "Probabilistic damage tolerance
analysis of aircraft structures", ASTM SiP

Pan, P.B. et al (1976): "Ultimate strength of tubular joints", Offshore Technology


Conference, OTC 2644, Houston, Texas.

Paris, P.C. andl Erdogan, F. (1963): "A Critical analysis of crack propagation laws",
J. of Basic Engineering, ASME, Vol. 85.

Payne, J.G. and Porter-Goff, R.F.D. (1986): "Experimental residual stress distributions
in welded tubular T nodes", Paper C134/86, Proc. Conf. on Fatigue and Crack Growth
in Offshore Structures", lMechE-86, London.

Plane, C.A. et al (1987): "The determination of safety factors for defect assessment
using reliability analysis methods", Proc. Conf. on Integrity of Offshore Structures,
IGS-87, Glasgow, U.K.

Porter-Goff, R.F.D. (1988): "Residual stresses in welded tubular nodes", Proc. Conf.
on Fatigue of Offshore Structures, (Eds) W.D. Dover and 0. Glinka, London, U.K.

Press, W.H. etal (1986): "Nwnerical Recipes",


Cambridge University Press

Rackwitz, R. and Fiessler,B. (1978): "Structural reliability under combined random


load sequences", Computers and Structures, Vol.9.

Ranganathan, R. and Deshpande, A.G. (1987): "Generation of dominant modes and


reliability analysis of frames", Structural Safety, Vol.4.

Rashedi, M.R. and Moses, F. (1988): "Identification of failure modes in system


reliability", ASCE, J. of Structural Engineering, Vol.114.
R-12
Reber, J.B. (1972): "Ultimate strength design for tubular joints", Offshore Technology
Conference, OTC 1664, Houston, Texas.

Reynolds, A.G. and Sharp, J.V. (1990): "The fatigue performance of tubular joints -
an overview of recent work to revise Department of Energy guidance", Conf. on
Integrity of Offshore Structures, Glasgow, U.K.

Rhee, H.C. (1985): "Comparative evaluation of fracture mechanics methodologies as


applied to offshore structural design and integrity analysis", J. of Energy Resources
Technology, Vol 107.

Rhee, H.C. and Kanninen, M. (1988): "Opportunities for application of fracture


mechanics for offshore structures", Applied Mechanics Reviews Vol 41, No.2.

Rice, S.O. (1944): "Mathematical analysis of random noise", Bell System Technical
Journal, Vols. 23, and 24.

Ritchie, D. et al (1987): "Stress intensity factor in an offshore tubular joint test


specimen", Proc. 3rd mt. symposium on Numerical Fracture Mechanics, San Antonio,
Texas.

Ritchie, D. (1988): "A fatigue crack growth model for thickness transition girth
welds", Conf. on Behaviour of Offshore Structures, Trondheim, Norway.

Rooke, D.P. and Cartwright, D.J. (1976): "Compediwn of stress intensity factors",
HMSO, London, U.K.

Sablok, A.K. and Hart, W.}I. (1990): "An experimental investigation of plate thickness
and weld profile effects pertaining to fatigue design of offshore structures", Proc.
Offshore Mechanics and Arctic Engineering Conference, Houston, Texas.

Sears, R.J. and Shetty, N.K. (1992): "RASOS evaluation study", WS Atkins Report
No. M1658/R01 for Amoco (UK) Exploration Company.

Sobzcyk, K. (1986): "Modelling of random fatigue crack growth", Engineering


Fracture Mechanics, Vol 24.

Sarpakaya, T. and Issacson, M. (1981): "Mechanics of wave forces on offshore


structures", Van Nostrand, New York, 1981.

Schinkowski, K. (1986): "NLPQL: A fortran subroutine solving constrained nonlinear


programming problems", Annals of Operations Research,VoL5_-

Scott, P.M. and Thorpe, T.W. (1981): "A critical review of crack tip stress intensity
factors for semi-elliptical surface cracks", Fatigue of Engineermg Materials and
Structures, Vol.4.
R-13
Shetty, N.K. and Baker, MJ. (1990a): "Fatigue reliability of tubular joints in offshore
structures: Fatigue loading", Proc. 9th Offshore Mechanics and Arctic Engineering
Conf., Houston, Texas.

Shetty, N.K. and Baker, Mi. (1990b): "Fatigue reliability of tubular joints in offshore
structures: Crack propagation model", Proc. 9th Offshore Mechanics and Arctic
Engineering Conf., Houston, Texas.

Shetty, N.K. and Baker, MJ. (1990c): "Fatigue reliability of tubular joints in offshore
structures: Reliability analysis", Proc. 9th Offshore Mechanics and Arctic Engineering
Conf., Houston, Texas.

Shetty, N.K. and Baker, MJ. (1991a): "Fatigue-fracture assessment of tubular joints",
RASOS_F Users Manual.

Shetty, N.K. and Baker, Mi. (1991b): "Fatigue-fracture assessment of tubular joints",
RASOS_F Programmers Manual.

Shetty, N.K. and Baker, Mi. (1991c): "Fatigue-fracture assessment of tubular joints",
RASOS_F Theoretical Manual.

Shetty, N.K. and Baker, Mi. (1991d): "Reliability analysis using failure-tree
enumeration", RASOS_R Users Manual

Shetty, N.K. et al (1991): "Reliability analysis using failure-tree enumeration",


RASOS_R Programmers Manual.

Skjong, R. and Torhaug, R. (1991): "Rational methods for fatigue design and
inspection planning of offshore structures", Marine Structures, Vol.4.

Snijder, RH. et al (1988), "Fatigue crack growth modelling for multiple initiated
cracks at the weld toes in tubular joints", Behaviour of offshore structures, BOSS-88,
Tmndheim, Norway.

Solin, J. (1990): "Analysis of spectrum fatigue tests in sea water", Conf. Offshore
Mechanics and Arctic Engineering, Vol.2.

Soreide, T.H. et al (1986): "Collapse analysis of framed offshore structures", Offshore


Technology Conference, OTC 5302, Houston, Texas.

Stahl, B. and Geyer, J.F. (1984): "Fatigue reliability of parallel member systems",
ASCE, Journal of Structural Engineering, Vol.110, No.10.

Stewart, G. and van de Grail, J.W. (1990): "A methodology for platform collapse
analysis based on linear superposition", Offshore Technology Conference, OTC 6311,
Houston, Texas.
R-14

Tada, H. et al (1973): 'The stress analysis of cracks handbook", Dell Research


Corporation, Hellertown, Pennysylvania.

Thoft-Christensen, P. and Baker, M.J. (1982): "Structural reliability theory and its
applications", Springer-Verlag, Berlin.

Thoft-Christensen, P. and Murotsu, Y. (1986): "Application of Structural System


Reliability Theory", Spriger-Verlag, Berlin.

Thorpe, T.W., and Sharp, J.V. (1990): "The effect of uncertainties in fatigue life on
design and inspection of offshore structures", Conf. on Integrity of Offshore
Structures, Glasgow, U.K

Turner, C.E., (1984): "Methods for post-yield fracture safety assessment", in "Post-
yield Fracture Mechanics", (eds. Latzko D.G.H. et al), 2nd Edition, Elsevier Applied
Science Publishers, London, U.K.

Turner, R.C. and Baker, MJ. (1987): "Structural system reliability analysis using
multi-dimensional limit-state criteria", Proc. First IFIP WG 7.5 Working Conference,
Aalborg, Denmark.

Turner, R.C. and Baker, MJ. (1988): "A probabilistic wave model for reliability
analysis", Proc. Second IFIP WG 7.5 Working Conference, London, U.K.

Turner, R.C. et al (1988): "The virtual distortion method applied to the reliability
analysis of offshore structures", Proc. Second IFIP WG 7.5 Working Conference,
London, U.K.

Turner, R.C. and Baker, M.J. (1988): "effect of wave phase angle on system reliability
of jacket structures", BRITE P 1270, Task 111-3, Internal Report.

Tvedt, L. (1988): "Second order reliability by an exact integral", Proc. 2nd IFIP
Working Conf. on Reliability and Optimization of Structural Systems, Springer.

Ueda, Y. and Rashed, S.M.H. (1991): "Modern method of ultimate strength analysis
of offshore structures", mt. Journal of Offshore and Polar Engineering, Vol.1, No.1.

Underwater Engineering Group, (1985): "Design of Tubular Joints for Offshore


Structures", UEG-UR33, London.

Virkier, D.A., Hillberry, B.M. and Goel, P.K. (1979): "The statistical nature of fatigue
crack propagation", J. Eng. Materials and Technology, Trans. ASME VoLlOL-

Watwood, V B. (1979): "Mechanism generation for hmit analysis of frames", ASCE,


Journal of the Structural Division, Vol.109, No.ST1.
R-15
Winterstein, S.R. and Haver, S. (1990): "Statistical uncertainties in wave heights and
combined loads on offshore structures", Proc. Offshore Mechanics and Arctic
Engineering Conference, Vol 2, Houston, Texas.

Wirsching, P.14. (1980): "Fatigue reliability of welded joints in offshore structures",


Proc. Offshore Technology Conference, Paper OTC 3380, Houston.

Wirsching, P.H. and Light, M.C. (1980): "Fatigue under wide-band random stresses",
J. of Structural Div., ASCE, VoL 106, No. ST7.

Wirsching, P.H. (1984): "Fatigue reliability for offshore structures", ASCE, J. of


Structural Engineering, Vol. 110, No. 10.

Wirsching, P.H., Ortiz, K. and Chen, Y.N. (1987): "Fracture mechanics fatigue model
in a reliability format", Proc. Offshore Mechanics and Arctic Engineering Conference,
Houston.

Wordsworth, AC. and Smedly, G.P. (1978): "Stress concentrations at un-stiffened


tubular joints", European Offshore Steels Research Seminar, Paper No. 26, Cambridge.

Wordsworth, A.C. (1981): "Stress concentration factors at K and KT tubular joints",


Coni. on Fatigue in Offshore Structural Steels, Institution of Civil Engineers, London.

Wylde, J.G. and McDonald, A. (1981): "Modes of fatigue crack development and
stiffness measurements in welded tubular joints", Paper 9, Conf. on Fatigue in
Offshore Structural Steels, ICE, London, U.K.

Yao, J.T.P. et al (1986): "Stochastic fatigue, fracture and damage analysis", Structural
Safety, Vol. 3.

Yura, J.A. et al (1980): "Ultimate capacity equations for tubular joints", Offshore
Technology Conference, 0TC3690, Houston, Texas.

Zhao, W. (1989): "Reliability analysis offatigue and fracture under random loading",
Ph.D. Thesis, University of London.

Thao, W. and Baker, Mi. (1990): "A new stress-range distribution model for fatigue
analysis under wave loading", in Environmental Forces on Offshore Structures and
Their Prediction, Kluwer.
Appendix A
OUTLINE OF THE RASOS SOFTWARE

Reliability Analysis System for Offshore Structures (RASOS) has been developed as
a part of the CEC sponsored BR1TE Project No. P 1270 and represents a joint effort
of 16 European companies and universities. It is one of the most powerful and
advanced computer software available for system reliability of jacket-type offshore
structures. RASOS is a suite of 11 closely interlinked software modules designed to
carry out both deterministic and probabilistic analysis of jacket structures with
minimum duplication of effort. Each of the modules feature a number of
STOP/RESTART options to enable the user to study the results of initial analyses
before going to the next stage. Main functions of each of the RASOS modules is
outlined in the following.

RASOS_M Model generation for truss and frame type structures, identification of
joint type based on geometry, input of stub and can properties for
joints, section and material properties and generation of gravity loading.

RASOS_E Environmental load generation due to the action of wave, wind and
current. Mean value of forces are computed in the form of distributed
loading on members or equivalent nodal loads. The 8 deterministic
components of environmental loading are computed and written to the
data base for later use in reliability analysis.

RASOS_L Stiffness generation and linear elastic analysis of jacket frames with
rigid or flexible joints. User defined static loads, self weight,
mean-value environmental loads and 8 deterministic components of
environmental loading are analyzed as separate load cases and
corresponding responses computed.

RASOSH Calculation of stress concentration factors for tubular joints. Spectral


analysis of the structure for a set of representative sea-states
A-2

considering wave directionality and directional spreading to compute


spectral moments of hot-spot stresses at various tubular joints. A
short-term density function model can be used for computing
equivalent stress ranges. Alternatively, the stress cycle contributions
from various sea-states can be combined and fitted into a single
long-term mixed Weibull density function.

RASOS_F Deterministic fatigue and fracture assessment of tubular joints. A


fracture mechanics fatigue crack propagation model is used to estimate
crack growth through the thickness and at the surface of a tubular joint
accounting explicitly for the effects of weld geometry, residual stress,
crack coalescence, load shedding etc. The failure of the damaged joint
by ductile fracture can be assessed using the CEGB R6 or the BS PD
6493 approaches. The program also offers plastic collapse assessment
and fatigue analysis using an S-N approach.

RASOSD Generation of structural influence matrix, input of non-linear member


properties and calculation of member-structure interaction springs for
use during deterministic collapse analysis or system reliability analysis.

RASOS_P Deterministic analysis of piles to determine foundation stiffness and


load bearing capacity for both axial and lateral loading. Probabilistic
assessment of the foundation system taking into account uncertain soil
properties and correlation between groups of piles.

RASOS_C Deterministic progressive collapse analysis based on the virtual


distortion method (VDM). Calculation of member capacities using
advanced 3-D non-linear buckling analysis including
softening/hardening behaviour and joint strength. Imperfections in the
form of initial out-of-straightness or dents can be considered. Global
boundary conditions allow non-linear support springs representing pile
foundation. A complete sequence of elemenLal failures is traced
without reformulation of the structural stiffness matrix.
A-3
RASOS_R System reliability analysis using failure-tree enumeration. A response
surface approach is used for environmental load modelling with 12
basic random variables. Virtual distortion method of non-linear
structural analysis allows treatment of elements with general softening!
hardening post-limit behaviour. A multi-constraint optimization
technique is used to search for the joint beta-point of a failure sequence
without straying into physically inadmissible regions of the probability
space. A total of 9 limit-states are available for describing component
resistance which includes 3-D elasto-plastic buckling model for tubular
members with imperfections and a tubular joint model with fatigue/
fracture effects. A total of 60 variable types, either independent or
correlated, can be used for uncertainty modelling.

RASOS_S System reliability analysis using advanced Monte Carlo simulation in


conjunction with efficient non-linear analysis using VDM. Structural
failure criteria can be in terms of critical displacements at selected
nodes or global collapse. The sampling efficiency of Monte Carlo
simulation is enhanced by exclusion of safe regions bounded by
hyperplanes obtained from FORM analysis. Techniques such as
importance sampling and directional sampling are also available.

RASOS_O Reliability based structural optimization of jackets in which the total


weight of the structure is minimized while maintaining specified
minimum levels of reliability for each element. Virtual distortion
method is used for efficient structural re-analyses simulating material
redistribution. Solution includes member diameters, thicknesses and
reliability indices for individual members and Level 1 system reliability
estimate for the optimized structure.

RDB All the modules use RASOS master control shell routine for the control
of the analysis sequence and the RASOS Data Base (RDB) for the
communication between individual modules.
B-i

Appendix B

LIMIT-STATE TYPES CONSIDERED IN RASOS

A. Member Limit-States:

1 - Plasticity at a member end using "plastic beam-column model"

2 - Strain-fracture at a member end using "plastic beam-column


model"

3 - Plasticity at a member end using "non-linear beam-column


model"

4 - Strain-fracture at a member end using "non-linear beam-column


model"

5-10 - Future use

B. Joint Limit-States:

11 - Fatigue failure using S-N approach

12 - Fatigue failure using fracture mechanics approach

13 - Fatigue followed by plastic collapse of the joint

14 - Fatigue followed by fracture of the joint


using C.E.G.B-R6 approach

15 - Fatigue followed by fracture of the joint


using B.S.P.D-6493 approach

16-20 - Future use


c-i
Appendix C

CODE NUMBERS FOR BASIC VARIABLES

A. User Defined Load Variables:

1-10 Corresponding to user defined load sets defined in RASOS_L

B. Extreme Environmental Loading Variables:

11 Wave height
12 Wave period
13 Water depth
14 Marine growth
15-20 Hydrodynamic and aerodynamic coefficients
21 Current speed
22 Wind speed
23-30 Future use

C. Fatigue Loading Variables:

31 Model uncertainty in response transfer function


32 Long-term Weibull model parameter WA
33 Long-term Weibull model parameter WB
34 Long-term Weibull model parameter WC
35 Long-term Weibull model parameter WD
36 Long-term Weibull model parameter WW
37 Long-term average total number of stress cycles
38-50 Future use

D. Member Resistance Variables:



71 Yield stress in tension
72 Yield stress in compression
73 Hardening slope in tension
74 Softening slope in compression
75 Strain-fracture limit in tension
76 Strain-fracture limit in compression
77 Initial imperfection in Y-direction
78 Initial imperfection in Z-direction
79 Uncertainty of "plastic beam-column model"
C-2

80 Uncertainty of "non-linear beam-column model"


81-90 Future use

E. Joint Resistance Variables:

91 Inverse slope of first part of S-N curve


92 Intercept of first part of S-N curve
93 Stress cut-off point for the first part
94 Inverse slope of second part of S-N curve
95 Intercept of second part of S-N curve
96 Stress cut-off point for the second part
97 Thickness correction exponent on stress
98 Miner's damage sum at failure
99-100 Future use

101 Exponent in Paris' law for crack growth (m)


102 Coefficient in Paris' law for crack growth (C)
103 Threshold stress intensity factor range
104 Tensile residual stress at the plate surface
105 Fracture toughness of the plate (K or CTOD)
106 Yield strength of the chord plate
107 Initial crack depth
108 Initial crack semi-length
109 Uncertainty in stress concentration factors
110 Model uncertainty in stress intensity factors
111 Model uncertainty in collapse load calculations
112-120 Future use
D-1
Appendix D
BRITE JACKET PLATFORM

The structure considered for reliability analysis examples presented in the thesis is a
6-legged fixed jacket platform standing in a water depth of 60m. The structural
configuration and member dimensions are based on a real platform for similar water
depth but the environmental conditions used in the examples is fictitious, designed to
demonstrate various aspects of probabilistic analysis. The platform is given a
fictitious name called "BRITE", the salient features of which are described below.

Structural Geometry:

The computer model of the jacket structure is shown in figures D.1-D.4. The structure
has 6-legs with piles going through them and to which they are connected by cement
grout. In the examples studied, however, the behaviour of pile foundation is not
considered and the jacket legs are assumed to be fixed at the sea-bed level. The
structure has 2 vertical frames parallel to X-axis, 2 vertical frames parallel to Y-axis
and 6 horizontal frames. The structure is symmetric about both X- and Y-axes. The
stiffness of the main horizontal members in the upper-most level are increased to
reflect the stiffness of the deck. The dimensions of members is given in Table D.1.
The stub and can sections for tubular joints have been deliberately omitted to study
the influence of fatigue deterioration on system reliability. Secondary members such
as conductors, risers, boat landing etc. have not been modelled and as a consequence
environmental loading will be some what underestimated.

All the horizontal members at the sea-bed level and the vertical members between this
level and the next upper level are called "Level 1" members, and similarly at higher
levels. The node numbers are marked on the figures. The element numbers are
derived by writing the end node numbers adjacent to each other, the lower node
number first, and dropping the zeros. For example, a member connecting nodes 203
and 306 is given an element number: 2(0)3 3(0)6 = 2336.
D-2

Table D.1: Dimensions of Tubular Members.

Diameter Thickness
Member Group
(m) (m)
Outer Jacket Legs (Lev- l/2/3,No: 12) 1.60 0.06
Middle Jacket Legs (Lev-1/2/3,No:6) 1.40 0.04
Outer Deck Legs (Lev-4/5,No: 8) 1.40 0.04
Middle Deck Legs (Lev-4/5,No:4) 0.80 0.03
Horizontal Braces (Lev-1i2/3/4,No:32) 1.60 0.06
Horizontal Braces (Lev-5/6,No: 16) 0.60 0.025
Horizontal Diagonals (Lev-1 to 6,No:30) 0.60 0.025
Vertical Diagonals (Lev- 1f2/3,No:30) 0.80 0.03
Vertical Diagonals (Lev-4/5,No: 16) 0.60 0.025

Gravity and Environmental Loading:

The self weight of the jacket structure from the members included in the computer
model are automatically computed by the RASOS program. The weight of the deck
(4500 tons) is applied as concentrated loads at the top-most leg nodes. The wind
loading on the deck is not considered but the wind loading on the exposed parts of the
jacket are automatically computed by the program. The parameters used in
deterministic analysis for computing design environmental loading are given in
Table D.2. The probabilistic modelling of basic variables is given in the main thesis.

Table D.2: Design environmental loading parameters.

Water Depth 60.Om


Wave Height 22.0 m
Wave Period 12.0 sec
Wave Direction Along +ve Y
Wind Speed 43.5 rn/sec
Current Speed 1.0 rn/sec
Drag Coefficient 0.7
Inertia Coefficient 1.8
--

I
I
I
N4
I
z

I
I
0

Das könnte Ihnen auch gefallen